FFTW

for version 3.3.3, 25 November 2012

Matteo Frigo Steven G. Johnson

This manual is for FFTW (version 3.3.3, 25 November 2012). Copyright c 2003 Matteo Frigo. Copyright c 2003 Massachusetts Institute of Technology. Permission is granted to make and distribute verbatim copies of this manual provided the copyright notice and this permission notice are preserved on all copies. Permission is granted to copy and distribute modified versions of this manual under the conditions for verbatim copying, provided that the entire resulting derived work is distributed under the terms of a permission notice identical to this one. Permission is granted to copy and distribute translations of this manual into another language, under the above conditions for modified versions, except that this permission notice may be stated in a translation approved by the Free Software Foundation.

Table of Contents
1 2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 Tutorial . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.1 2.2 2.3 2.4 2.5 Complex One-Dimensional DFTs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 Complex Multi-Dimensional DFTs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 One-Dimensional DFTs of Real Data . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 Multi-Dimensional DFTs of Real Data . . . . . . . . . . . . . . . . . . . . . . . . . . 7 More DFTs of Real Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 2.5.1 The Halfcomplex-format DFT . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 2.5.2 Real even/odd DFTs (cosine/sine transforms) . . . . . . . . . . . . 11 2.5.3 The Discrete Hartley Transform . . . . . . . . . . . . . . . . . . . . . . . . . . 13

Other Important Topics . . . . . . . . . . . . . . . . . . . . . . . 15

3.1 3.2 SIMD alignment and fftw malloc . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Multi-dimensional Array Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2.1 Row-major Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2.2 Column-major Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2.3 Fixed-size Arrays in C . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2.4 Dynamic Arrays in C . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2.5 Dynamic Arrays in CThe Wrong Way . . . . . . . . . . . . . . . . . . 3.3 Words of WisdomSaving Plans . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.4 Caveats in Using Wisdom . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 15 15 16 16 17 17 18 18

FFTW Reference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
4.1 Data Types and Files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.1.1 Complex numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.1.2 Precision . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.1.3 Memory Allocation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2 Using Plans . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3 Basic Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3.1 Complex DFTs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3.2 Planner Flags . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3.3 Real-data DFTs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3.4 Real-data DFT Array Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3.5 Real-to-Real Transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3.6 Real-to-Real Transform Kinds . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.4 Advanced Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.4.1 Advanced Complex DFTs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.4.2 Advanced Real-data DFTs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.4.3 Advanced Real-to-real Transforms . . . . . . . . . . . . . . . . . . . . . . . . 4.5 Guru Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.5.1 Interleaved and split arrays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 21 21 22 22 24 24 25 27 28 29 30 31 31 33 33 34 34

ii 4.5.2 Guru vector and transform sizes . . . . . . . . . . . . . . . . . . . . . . . . . . 4.5.3 Guru Complex DFTs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.5.4 Guru Real-data DFTs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.5.5 Guru Real-to-real Transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.5.6 64-bit Guru Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.6 New-array Execute Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.7 Wisdom . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.7.1 Wisdom Export . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.7.2 Wisdom Import . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.7.3 Forgetting Wisdom . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.7.4 Wisdom Utilities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.8 What FFTW Really Computes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.8.1 The 1d Discrete Fourier Transform (DFT) . . . . . . . . . . . . . . . . 4.8.2 The 1d Real-data DFT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.8.3 1d Real-even DFTs (DCTs) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.8.4 1d Real-odd DFTs (DSTs) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.8.5 1d Discrete Hartley Transforms (DHTs) . . . . . . . . . . . . . . . . . . 4.8.6 Multi-dimensional Transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34 35 36 37 37 38 40 40 40 41 41 42 42 42 43 44 45 46

FFTW 3.3.3

Multi-threaded FFTW . . . . . . . . . . . . . . . . . . . . . . . . 49
5.1 5.2 5.3 5.4 Installation and Supported Hardware/Software . . . . . . . . . . . . . . . . Usage of Multi-threaded FFTW . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . How Many Threads to Use? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Thread safety . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49 49 51 51

Distributed-memory FFTW with MPI . . . . . . 53

6.1 6.2 6.3 6.4 FFTW MPI Installation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Linking and Initializing MPI FFTW . . . . . . . . . . . . . . . . . . . . . . . . . . . 2d MPI example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . MPI Data Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.4.1 Basic and advanced distribution interfaces . . . . . . . . . . . . . . . . 6.4.2 Load balancing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.4.3 Transposed distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.4.4 One-dimensional distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.5 Multi-dimensional MPI DFTs of Real Data . . . . . . . . . . . . . . . . . . . . 6.6 Other multi-dimensional Real-Data MPI Transforms . . . . . . . . . . . 6.7 FFTW MPI Transposes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.7.1 Basic distributed-transpose interface . . . . . . . . . . . . . . . . . . . . . . 6.7.2 Advanced distributed-transpose interface . . . . . . . . . . . . . . . . . 6.7.3 An improved replacement for MPI Alltoall . . . . . . . . . . . . . . . 6.8 FFTW MPI Wisdom . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.9 Avoiding MPI Deadlocks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.10 FFTW MPI Performance Tips . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.11 Combining MPI and Threads . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.12 FFTW MPI Reference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.12.1 MPI Files and Data Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.12.2 MPI Initialization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.12.3 Using MPI Plans . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53 54 54 56 56 58 58 59 60 62 62 63 63 64 64 65 66 66 67 67 68 68

iii 6.12.4 MPI Data Distribution Functions . . . . . . . . . . . . . . . . . . . . . . . . 6.12.5 MPI Plan Creation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.12.6 MPI Wisdom Communication . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.13 FFTW MPI Fortran Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69 70 74 74

Calling FFTW from Modern Fortran . . . . . . . . 77

7.1 Overview of Fortran interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.1.1 Extended and quadruple precision in Fortran . . . . . . . . . . . . . 7.2 Reversing array dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.3 FFTW Fortran type reference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.4 Plan execution in Fortran . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.5 Allocating aligned memory in Fortran . . . . . . . . . . . . . . . . . . . . . . . . . 7.6 Accessing the wisdom API from Fortran . . . . . . . . . . . . . . . . . . . . . . . 7.6.1 Wisdom File Export/Import from Fortran . . . . . . . . . . . . . . . . 7.6.2 Wisdom String Export/Import from Fortran . . . . . . . . . . . . . . 7.6.3 Wisdom Generic Export/Import from Fortran . . . . . . . . . . . . 7.7 Defining an FFTW module . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77 78 78 80 81 82 83 83 84 84 85

Calling FFTW from Legacy Fortran . . . . . . . . . 87

8.1 8.2 8.3 8.4 8.5 Fortran-interface routines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . FFTW Constants in Fortran . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . FFTW Execution in Fortran . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Fortran Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Wisdom of Fortran? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87 88 88 89 91

9 10

Upgrading from FFTW version 2 . . . . . . . . . . . . 93 Installation and Customization . . . . . . . . . . . . . 97

Installation on Unix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97 Installation on non-Unix systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100 Cycle Counters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100 Generating your own code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101

10.1 10.2 10.3 10.4

11 12 13 14

Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . 103 License and Copyright . . . . . . . . . . . . . . . . . . . . . . 105 Concept Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107 Library Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109

Chapter 1: Introduction

1 Introduction
This manual documents version 3.3.3 of FFTW, the Fastest Fourier Transform in the West. FFTW is a comprehensive collection of fast C routines for computing the discrete Fourier transform (DFT) and various special cases thereof. FFTW computes the DFT of complex data, real data, even- or odd-symmetric real data (these symmetric transforms are usually known as the discrete cosine or sine transform, respectively), and the discrete Hartley transform (DHT) of real data. The input data can have arbitrary length. FFTW employs O(n log n) algorithms for all lengths, including prime numbers. FFTW supports arbitrary multi-dimensional data. FFTW supports the SSE, SSE2, AVX, Altivec, and MIPS PS instruction sets. FFTW includes parallel (multi-threaded) transforms for shared-memory systems. Starting with version 3.3, FFTW includes distributed-memory parallel transforms using MPI. We assume herein that you are familiar with the properties and uses of the DFT that are relevant to your application. Otherwise, see e.g. The Fast Fourier Transform and Its Applications by E. O. Brigham (Prentice-Hall, Englewood Cliffs, NJ, 1988). Our web page also has links to FFT-related information online. In order to use FFTW effectively, you need to learn one basic concept of FFTWs internal structure: FFTW does not use a fixed algorithm for computing the transform, but instead it adapts the DFT algorithm to details of the underlying hardware in order to maximize performance. Hence, the computation of the transform is split into two phases. First, FFTWs planner learns the fastest way to compute the transform on your machine. The planner produces a data structure called a plan that contains this information. Subsequently, the plan is executed to transform the array of input data as dictated by the plan. The plan can be reused as many times as needed. In typical high-performance applications, many transforms of the same size are computed and, consequently, a relatively expensive initialization of this sort is acceptable. On the other hand, if you need a single transform of a given size, the one-time cost of the planner becomes significant. For this case, FFTW provides fast planners based on heuristics or on previously computed plans. FFTW supports transforms of data with arbitrary length, rank, multiplicity, and a general memory layout. In simple cases, however, this generality may be unnecessary and confusing. Consequently, we organized the interface to FFTW into three levels of increasing generality. The basic interface computes a single transform of contiguous data. The advanced interface computes transforms of multiple or strided arrays. The guru interface supports the most general data layouts, multiplicities, and strides. We expect that most users will be best served by the basic interface, whereas the guru interface requires careful attention to the documentation to avoid problems. Besides the automatic performance adaptation performed by the planner, it is also possible for advanced users to customize FFTW manually. For example, if code space is a concern, we

FFTW 3.3.3

provide a tool that links only the subset of FFTW needed by your application. Conversely, you may need to extend FFTW because the standard distribution is not sufficient for your needs. For example, the standard FFTW distribution works most efficiently for arrays whose size can be factored into small primes (2, 3, 5, and 7), and otherwise it uses a slower general-purpose routine. If you need efficient transforms of other sizes, you can use FFTWs code generator, which produces fast C programs (codelets) for any particular array size you may care about. For example, if you need transforms of size 513 = 19 33 , you can customize FFTW to support the factor 19 efficiently. For more information regarding FFTW, see the paper, The Design and Implementation of FFTW3, by M. Frigo and S. G. Johnson, which was an invited paper in Proc. IEEE 93 (2), p. 216 (2005). The code generator is described in the paper A fast Fourier transform compiler, by M. Frigo, in the Proceedings of the 1999 ACM SIGPLAN Conference on Programming Language Design and Implementation (PLDI), Atlanta, Georgia, May 1999. These papers, along with the latest version of FFTW, the FAQ, benchmarks, and other links, are available at the FFTW home page. The current version of FFTW incorporates many good ideas from the past thirty years of FFT literature. In one way or another, FFTW uses the Cooley-Tukey algorithm, the prime factor algorithm, Raders algorithm for prime sizes, and a split-radix algorithm (with a conjugate-pair variation pointed out to us by Dan Bernstein). FFTWs code generator also produces new algorithms that we do not completely understand. The reader is referred to the cited papers for the appropriate references. The rest of this manual is organized as follows. We first discuss the sequential (singleprocessor) implementation. We start by describing the basic interface/features of FFTW in Chapter 2 [Tutorial], page 3. Next, Chapter 3 [Other Important Topics], page 15 discusses data alignment (see Section 3.1 [SIMD alignment and fftw malloc], page 15), the storage scheme of multi-dimensional arrays (see Section 3.2 [Multi-dimensional Array Format], page 15), and FFTWs mechanism for storing plans on disk (see Section 3.3 [Words of Wisdom-Saving Plans], page 18). Next, Chapter 4 [FFTW Reference], page 21 provides comprehensive documentation of all FFTWs features. Parallel transforms are discussed in their own chapters: Chapter 5 [Multi-threaded FFTW], page 49 and Chapter 6 [Distributedmemory FFTW with MPI], page 53. Fortran programmers can also use FFTW, as described in Chapter 8 [Calling FFTW from Legacy Fortran], page 87 and Chapter 7 [Calling FFTW from Modern Fortran], page 77. Chapter 10 [Installation and Customization], page 97 explains how to install FFTW in your computer system and how to adapt FFTW to your needs. License and copyright information is given in Chapter 12 [License and Copyright], page 105. Finally, we thank all the people who helped us in Chapter 11 [Acknowledgments], page 103.

Chapter 2: Tutorial

2 Tutorial
This chapter describes the basic usage of FFTW, i.e., how to compute the Fourier transform of a single array. This chapter tells the truth, but not the whole truth. Specifically, FFTW implements additional routines and flags that are not documented here, although in many cases we try to indicate where added capabilities exist. For more complete information, see Chapter 4 [FFTW Reference], page 21. (Note that you need to compile and install FFTW before you can use it in a program. For the details of the installation, see Chapter 10 [Installation and Customization], page 97.) We recommend that you read this tutorial in order.1 At the least, read the first section (see Section 2.1 [Complex One-Dimensional DFTs], page 3) before reading any of the others, even if your main interest lies in one of the other transform types. Users of FFTW version 2 and earlier may also want to read Chapter 9 [Upgrading from FFTW version 2], page 93.

2.1 Complex One-Dimensional DFTs

Plan: To bother about the best method of accomplishing an accidental result. [Ambrose Bierce, The Enlarged Devils Dictionary.] The basic usage of FFTW to compute a one-dimensional DFT of size N is simple, and it typically looks something like this code: #include <fftw3.h> ... { fftw_complex *in, *out; fftw_plan p; ... in = (fftw_complex*) fftw_malloc(sizeof(fftw_complex) * N); out = (fftw_complex*) fftw_malloc(sizeof(fftw_complex) * N); p = fftw_plan_dft_1d(N, in, out, FFTW_FORWARD, FFTW_ESTIMATE); ... fftw_execute(p); /* repeat as needed */ ... fftw_destroy_plan(p); fftw_free(in); fftw_free(out); } You must link this code with the fftw3 library. On Unix systems, link with -lfftw3 -lm. The example code first allocates the input and output arrays. You can allocate them in any way that you like, but we recommend using fftw_malloc, which behaves like malloc except that it properly aligns the array when SIMD instructions (such as SSE and Altivec) are available (see Section 3.1 [SIMD alignment and fftw malloc], page 15). [Alternatively, we provide a convenient wrapper function fftw_alloc_complex(N) which has the same effect.]
1

You can read the tutorial in bit-reversed order after computing your first transform.

FFTW 3.3.3

The data is an array of type fftw_complex, which is by default a double[2] composed of the real (in[i][0]) and imaginary (in[i][1]) parts of a complex number. The next step is to create a plan, which is an object that contains all the data that FFTW needs to compute the FFT. This function creates the plan: fftw_plan fftw_plan_dft_1d(int n, fftw_complex *in, fftw_complex *out, int sign, unsigned flags); The first argument, n, is the size of the transform you are trying to compute. The size n can be any positive integer, but sizes that are products of small factors are transformed most efficiently (although prime sizes still use an O(n log n) algorithm). The next two arguments are pointers to the input and output arrays of the transform. These pointers can be equal, indicating an in-place transform. The fourth argument, sign, can be either FFTW_FORWARD (-1) or FFTW_BACKWARD (+1), and indicates the direction of the transform you are interested in; technically, it is the sign of the exponent in the transform. The flags argument is usually either FFTW_MEASURE or FFTW_ESTIMATE. FFTW_MEASURE instructs FFTW to run and measure the execution time of several FFTs in order to find the best way to compute the transform of size n. This process takes some time (usually a few seconds), depending on your machine and on the size of the transform. FFTW_ESTIMATE, on the contrary, does not run any computation and just builds a reasonable plan that is probably sub-optimal. In short, if your program performs many transforms of the same size and initialization time is not important, use FFTW_MEASURE; otherwise use the estimate. You must create the plan before initializing the input, because FFTW_MEASURE overwrites the in/out arrays. (Technically, FFTW_ESTIMATE does not touch your arrays, but you should always create plans first just to be sure.) Once the plan has been created, you can use it as many times as you like for transforms on the specified in/out arrays, computing the actual transforms via fftw_execute(plan): void fftw_execute(const fftw_plan plan); The DFT results are stored in-order in the array out, with the zero-frequency (DC) component in out[0]. If in != out, the transform is out-of-place and the input array in is not modified. Otherwise, the input array is overwritten with the transform. If you want to transform a different array of the same size, you can create a new plan with fftw_plan_dft_1d and FFTW automatically reuses the information from the previous plan, if possible. Alternatively, with the guru interface you can apply a given plan to a different array, if you are careful. See Chapter 4 [FFTW Reference], page 21. When you are done with the plan, you deallocate it by calling fftw_destroy_plan(plan): void fftw_destroy_plan(fftw_plan plan); If you allocate an array with fftw_malloc() you must deallocate it with fftw_free(). Do not use free() or, heaven forbid, delete. FFTW computes an unnormalized DFT. Thus, computing a forward followed by a backward transform (or vice versa) results in the original array scaled by n. For the definition of the DFT, see Section 4.8 [What FFTW Really Computes], page 42.

Chapter 2: Tutorial

If you have a C compiler, such as gcc, that supports the C99 standard, and you #include <complex.h> before <fftw3.h>, then fftw_complex is the native double-precision complex type and you can manipulate it with ordinary arithmetic. Otherwise, FFTW defines its own complex type, which is bit-compatible with the C99 complex type. See Section 4.1.1 [Complex numbers], page 21. (The C++ <complex> template class may also be usable via a typecast.) To use single or long-double precision versions of FFTW, replace the fftw_ prefix by fftwf_ or fftwl_ and link with -lfftw3f or -lfftw3l, but use the same <fftw3.h> header file. Many more flags exist besides FFTW_MEASURE and FFTW_ESTIMATE. For example, use FFTW_ PATIENT if youre willing to wait even longer for a possibly even faster plan (see Chapter 4 [FFTW Reference], page 21). You can also save plans for future use, as described by Section 3.3 [Words of Wisdom-Saving Plans], page 18.

2.2 Complex Multi-Dimensional DFTs

Multi-dimensional transforms work much the same way as one-dimensional transforms: you allocate arrays of fftw_complex (preferably using fftw_malloc), create an fftw_plan, execute it as many times as you want with fftw_execute(plan), and clean up with fftw_ destroy_plan(plan) (and fftw_free). FFTW provides two routines for creating plans for 2d and 3d transforms, and one routine for creating plans of arbitrary dimensionality. The 2d and 3d routines have the following signature: fftw_plan fftw_plan_dft_2d(int n0, int n1, fftw_complex *in, fftw_complex *out, int sign, unsigned flags); fftw_plan fftw_plan_dft_3d(int n0, int n1, int n2, fftw_complex *in, fftw_complex *out, int sign, unsigned flags); These routines create plans for n0 by n1 two-dimensional (2d) transforms and n0 by n1 by n2 3d transforms, respectively. All of these transforms operate on contiguous arrays in the C-standard row-major order, so that the last dimension has the fastest-varying index in the array. This layout is described further in Section 3.2 [Multi-dimensional Array Format], page 15. FFTW can also compute transforms of higher dimensionality. In order to avoid confusion between the various meanings of the the word dimension, we use the term rank to denote the number of independent indices in an array.2 For example, we say that a 2d transform has rank 2, a 3d transform has rank 3, and so on. You can plan transforms of arbitrary rank by means of the following function: fftw_plan fftw_plan_dft(int rank, const int *n, fftw_complex *in, fftw_complex *out, int sign, unsigned flags);
2

The term rank is commonly used in the APL, FORTRAN, and Common Lisp traditions, although it is not so common in the C world.

FFTW 3.3.3

Here, n is a pointer to an array n[rank] denoting an n[0] by n[1] by . . . by n[rank-1] transform. Thus, for example, the call fftw_plan_dft_2d(n0, n1, in, out, sign, flags); is equivalent to the following code fragment: int n[2]; n[0] = n0; n[1] = n1; fftw_plan_dft(2, n, in, out, sign, flags); fftw_plan_dft is not restricted to 2d and 3d transforms, however, but it can plan transforms of arbitrary rank. You may have noticed that all the planner routines described so far have overlapping functionality. For example, you can plan a 1d or 2d transform by using fftw_plan_dft with a rank of 1 or 2, or even by calling fftw_plan_dft_3d with n0 and/or n1 equal to 1 (with no loss in efficiency). This pattern continues, and FFTWs planning routines in general form a partial order, sequences of interfaces with strictly increasing generality but correspondingly greater complexity. fftw_plan_dft is the most general complex-DFT routine that we describe in this tutorial, but there are also the advanced and guru interfaces, which allow one to efficiently combine multiple/strided transforms into a single FFTW plan, transform a subset of a larger multidimensional array, and/or to handle more general complex-number formats. For more information, see Chapter 4 [FFTW Reference], page 21.

2.3 One-Dimensional DFTs of Real Data

In many practical applications, the input data in[i] are purely real numbers, in which case the DFT output satisfies the Hermitian redundancy: out[i] is the conjugate of out[ni]. It is possible to take advantage of these circumstances in order to achieve roughly a factor of two improvement in both speed and memory usage. In exchange for these speed and space advantages, the user sacrifices some of the simplicity of FFTWs complex transforms. First of all, the input and output arrays are of different sizes and types : the input is n real numbers, while the output is n/2+1 complex numbers (the non-redundant outputs); this also requires slight padding of the input array for in-place transforms. Second, the inverse transform (complex to real) has the side-effect of overwriting its input array, by default. Neither of these inconveniences should pose a serious problem for users, but it is important to be aware of them. The routines to perform real-data transforms are almost the same as those for complex transforms: you allocate arrays of double and/or fftw_complex (preferably using fftw_ malloc or fftw_alloc_complex), create an fftw_plan, execute it as many times as you want with fftw_execute(plan), and clean up with fftw_destroy_plan(plan) (and fftw_ free). The only differences are that the input (or output) is of type double and there are new routines to create the plan. In one dimension: fftw_plan fftw_plan_dft_r2c_1d(int n, double *in, fftw_complex *out, unsigned flags);

Chapter 2: Tutorial

fftw_plan fftw_plan_dft_c2r_1d(int n, fftw_complex *in, double *out, unsigned flags); for the real input to complex-Hermitian output (r2c ) and complex-Hermitian input to real output (c2r ) transforms. Unlike the complex DFT planner, there is no sign argument. Instead, r2c DFTs are always FFTW_FORWARD and c2r DFTs are always FFTW_BACKWARD. (For single/long-double precision fftwf and fftwl, double should be replaced by float and long double, respectively.) Here, n is the logical size of the DFT, not necessarily the physical size of the array. In particular, the real (double) array has n elements, while the complex (fftw_complex) array has n/2+1 elements (where the division is rounded down). For an in-place transform, in and out are aliased to the same array, which must be big enough to hold both; so, the real array would actually have 2*(n/2+1) elements, where the elements beyond the first n are unused padding. (Note that this is very different from the concept of zero-padding a transform to a larger length, which changes the logical size of the DFT by actually adding new input data.) The k th element of the complex array is exactly the same as the k th element of the corresponding complex DFT. All positive n are supported; products of small factors are most efficient, but an O(n log n) algorithm is used even for prime sizes. As noted above, the c2r transform destroys its input array even for out-of-place transforms. This can be prevented, if necessary, by including FFTW_PRESERVE_INPUT in the flags, with unfortunately some sacrifice in performance. This flag is also not currently supported for multi-dimensional real DFTs (next section). Readers familiar with DFTs of real data will recall that the 0th (the DC) and n/2-th (the Nyquist frequency, when n is even) elements of the complex output are purely real. Some implementations therefore store the Nyquist element where the DC imaginary part would go, in order to make the input and output arrays the same size. Such packing, however, does not generalize well to multi-dimensional transforms, and the space savings are miniscule in any case; FFTW does not support it. An alternative interface for one-dimensional r2c and c2r DFTs can be found in the r2r interface (see Section 2.5.1 [The Halfcomplex-format DFT], page 11), with halfcomplexformat output that is the same size (and type) as the input array. That interface, although it is not very useful for multi-dimensional transforms, may sometimes yield better performance.

2.4 Multi-Dimensional DFTs of Real Data

Multi-dimensional DFTs of real data use the following planner routines: fftw_plan fftw_plan_dft_r2c_2d(int n0, int n1, double *in, fftw_complex *out, unsigned flags); fftw_plan fftw_plan_dft_r2c_3d(int n0, int n1, int n2, double *in, fftw_complex *out, unsigned flags); fftw_plan fftw_plan_dft_r2c(int rank, const int *n, double *in, fftw_complex *out, unsigned flags);

FFTW 3.3.3

as well as the corresponding c2r routines with the input/output types swapped. These routines work similarly to their complex analogues, except for the fact that here the complex output array is cut roughly in half and the real array requires padding for in-place transforms (as in 1d, above).

As before, n is the logical size of the array, and the consequences of this on the the format of the complex arrays deserve careful attention. Suppose that the real data has dimensions n0 n1 n2 nd1 (in row-major order). Then, after an r2c transform, the output is an n0 n1 n2 (nd1 /2 + 1) array of fftw_complex values in row-major order, corresponding to slightly over half of the output of the corresponding complex DFT. (The division is rounded down.) The ordering of the data is otherwise exactly the same as in the complex-DFT case.

For out-of-place transforms, this is the end of the story: the real data is stored as a rowmajor array of size n0 n1 n2 nd1 and the complex data is stored as a row-major array of size n0 n1 n2 (nd1 /2 + 1) .

For in-place transforms, however, extra padding of the real-data array is necessary because the complex array is larger than the real array, and the two arrays share the same memory locations. Thus, for in-place transforms, the final dimension of the real-data array must be padded with extra values to accommodate the size of the complex datatwo values if the last dimension is even and one if it is odd. That is, the last dimension of the real data must physically contain 2(nd1 /2 + 1) double values (exactly enough to hold the complex data). This physical array size does not, however, change the logical array sizeonly nd1 values are actually stored in the last dimension, and nd1 is the last dimension passed to the plan-creation routine.

For example, consider the transform of a two-dimensional real array of size n0 by n1. The output of the r2c transform is a two-dimensional complex array of size n0 by n1/2+1, where the y dimension has been cut nearly in half because of redundancies in the output. Because fftw_complex is twice the size of double, the output array is slightly bigger than the input array. Thus, if we want to compute the transform in place, we must pad the input array so that it is of size n0 by 2*(n1/2+1). If n1 is even, then there are two padding elements at the end of each row (which need not be initialized, as they are only used for output).

Chapter 2: Tutorial

ny
0 ny1 1 ny+1 2 3 ny4 ny3 ny2 ny1

input, outofplace

0 0 ny

...

nx

nx1

ny/2+1
0 0 0 ny/2+1 1 ny/22 ny/21 ny/2 ny/2

...

output

nx

nx1

= double

= fftw_complex

ny + 2ny%2 = 2*(ny/2+1)
0 0 0 1 2 3 ny4 ny3 ny2 ny1 ny ny+1 ny+1

input, inplace

ny+2 ny+3

...

nx

nx1

Figure 2.1: Illustration of the data layout for a 2d nx by ny real-to-complex transform. Figure 2.1 depicts the input and output arrays just described, for both the out-of-place and in-place transforms (with the arrows indicating consecutive memory locations): These transforms are unnormalized, so an r2c followed by a c2r transform (or vice versa) will result in the original data scaled by the number of real data elementsthat is, the product of the (logical) dimensions of the real data. (Because the last dimension is treated specially, if it is equal to 1 the transform is not equivalent to a lower-dimensional r2c/c2r transform. In that case, the last complex dimension also has size 1 (=1/2+1), and no advantage is gained over the complex transforms.)

(padding)

10

FFTW 3.3.3

2.5 More DFTs of Real Data

FFTW supports several other transform types via a unified r2r (real-to-real) interface, so called because it takes a real (double) array and outputs a real array of the same size. These r2r transforms currently fall into three categories: DFTs of real input and complexHermitian output in halfcomplex format, DFTs of real input with even/odd symmetry (a.k.a. discrete cosine/sine transforms, DCTs/DSTs), and discrete Hartley transforms (DHTs), all described in more detail by the following sections. The r2r transforms follow the by now familiar interface of creating an fftw_plan, executing it with fftw_execute(plan), and destroying it with fftw_destroy_plan(plan). Furthermore, all r2r transforms share the same planner interface: fftw_plan fftw_plan_r2r_1d(int n, double *in, double *out, fftw_r2r_kind kind, unsigned flags); fftw_plan fftw_plan_r2r_2d(int n0, int n1, double *in, double *out, fftw_r2r_kind kind0, fftw_r2r_kind kind1, unsigned flags); fftw_plan fftw_plan_r2r_3d(int n0, int n1, int n2, double *in, double *out, fftw_r2r_kind kind0, fftw_r2r_kind kind1, fftw_r2r_kind kind2, unsigned flags); fftw_plan fftw_plan_r2r(int rank, const int *n, double *in, double *out, const fftw_r2r_kind *kind, unsigned flags); Just as for the complex DFT, these plan 1d/2d/3d/multi-dimensional transforms for contiguous arrays in row-major order, transforming (real) input to output of the same size, where n specifies the physical dimensions of the arrays. All positive n are supported (with the exception of n=1 for the FFTW_REDFT00 kind, noted in the real-even subsection below); products of small factors are most efficient (factorizing n-1 and n+1 for FFTW_REDFT00 and FFTW_RODFT00 kinds, described below), but an O(n log n) algorithm is used even for prime sizes. Each dimension has a kind parameter, of type fftw_r2r_kind, specifying the kind of r2r transform to be used for that dimension. (In the case of fftw_plan_r2r, this is an array kind[rank] where kind[i] is the transform kind for the dimension n[i].) The kind can be one of a set of predefined constants, defined in the following subsections. In other words, FFTW computes the separable product of the specified r2r transforms over each dimension, which can be used e.g. for partial differential equations with mixed boundary conditions. (For some r2r kinds, notably the halfcomplex DFT and the DHT, such a separable product is somewhat problematic in more than one dimension, however, as is described below.) In the current version of FFTW, all r2r transforms except for the halfcomplex type are computed via pre- or post-processing of halfcomplex transforms, and they are therefore not as fast as they could be. Since most other general DCT/DST codes employ a similar algorithm, however, FFTWs implementation should provide at least competitive performance.

Chapter 2: Tutorial

11

2.5.1 The Halfcomplex-format DFT

An r2r kind of FFTW_R2HC (r2hc ) corresponds to an r2c DFT (see Section 2.3 [One-Dimensional DFTs of Real Data], page 6) but with halfcomplex format output, and may sometimes be faster and/or more convenient than the latter. The inverse hc2r transform is of kind FFTW_HC2R. This consists of the non-redundant half of the complex output for a 1d real-input DFT of size n, stored as a sequence of n real numbers (double) in the format: r0 , r1 , r2 , . . . , rn/2 , i(n+1)/21 , . . . , i2 , i1 Here, rk is the real part of the k th output, and ik is the imaginary part. (Division by 2 is rounded down.) For a halfcomplex array hc[n], the k th component thus has its real part in hc[k] and its imaginary part in hc[n-k], with the exception of k == 0 or n/2 (the latter only if n is even)in these two cases, the imaginary part is zero due to symmetries of the real-input DFT, and is not stored. Thus, the r2hc transform of n real values is a halfcomplex array of length n, and vice versa for hc2r. Aside from the differing format, the output of FFTW_R2HC/FFTW_HC2R is otherwise exactly the same as for the corresponding 1d r2c/c2r transform (i.e. FFTW_FORWARD/FFTW_BACKWARD transforms, respectively). Recall that these transforms are unnormalized, so r2hc followed by hc2r will result in the original data multiplied by n. Furthermore, like the c2r transform, an out-of-place hc2r transform will destroy its input array. Although these halfcomplex transforms can be used with the multi-dimensional r2r interface, the interpretation of such a separable product of transforms along each dimension is problematic. For example, consider a two-dimensional n0 by n1, r2hc by r2hc transform planned by fftw_plan_r2r_2d(n0, n1, in, out, FFTW_R2HC, FFTW_R2HC, FFTW_MEASURE). Conceptually, FFTW first transforms the rows (of size n1) to produce halfcomplex rows, and then transforms the columns (of size n0). Half of these column transforms, however, are of imaginary parts, and should therefore be multiplied by i and combined with the r2hc transforms of the real columns to produce the 2d DFT amplitudes; FFTWs r2r transform does not perform this combination for you. Thus, if a multi-dimensional real-input/output DFT is required, we recommend using the ordinary r2c/c2r interface (see Section 2.4 [MultiDimensional DFTs of Real Data], page 7).

2.5.2 Real even/odd DFTs (cosine/sine transforms)

The Fourier transform of a real-even function f (x) = f (x) is real-even, and i times the Fourier transform of a real-odd function f (x) = f (x) is real-odd. Similar results hold for a discrete Fourier transform, and thus for these symmetries the need for complex inputs/outputs is entirely eliminated. Moreover, one gains a factor of two in speed/space from the fact that the data are real, and an additional factor of two from the even/odd symmetry: only the non-redundant (first) half of the array need be stored. The result is the real-even DFT (REDFT ) and the real-odd DFT (RODFT ), also known as the discrete cosine and sine transforms (DCT and DST ), respectively. (In this section, we describe the 1d transforms; multi-dimensional transforms are just a separable product of these transforms operating along each dimension.)

12

FFTW 3.3.3

Because of the discrete sampling, one has an additional choice: is the data even/odd around a sampling point, or around the point halfway between two samples? The latter corresponds to shifting the samples by half an interval, and gives rise to several transform variants denoted by REDFTab and RODFTab: a and b are 0 or 1, and indicate whether the input (a) and/or output (b) are shifted by half a sample (1 means it is shifted). These are also known as types I-IV of the DCT and DST, and all four types are supported by FFTWs r2r interface.3 The r2r kinds for the various REDFT and RODFT types supported by FFTW, along with the boundary conditions at both ends of the input array (n real numbers in[j=0..n-1]), are: FFTW_REDFT00 (DCT-I): even around j = 0 and even around j = n 1. FFTW_REDFT10 (DCT-II, the DCT): even around j = 0.5 and even around j = n 0.5. FFTW_REDFT01 (DCT-III, the IDCT): even around j = 0 and odd around j = n. FFTW_REDFT11 (DCT-IV): even around j = 0.5 and odd around j = n 0.5. FFTW_RODFT00 (DST-I): odd around j = 1 and odd around j = n. FFTW_RODFT10 (DST-II): odd around j = 0.5 and odd around j = n 0.5. FFTW_RODFT01 (DST-III): odd around j = 1 and even around j = n 1. FFTW_RODFT11 (DST-IV): odd around j = 0.5 and even around j = n 0.5. Note that these symmetries apply to the logical array being transformed; there are no constraints on your physical input data. So, for example, if you specify a size-5 REDFT00 (DCT-I) of the data abcde, it corresponds to the DFT of the logical even array abcdedcb of size 8. A size-4 REDFT10 (DCT-II) of the data abcd corresponds to the size-8 logical DFT of the even array abcddcba, shifted by half a sample. All of these transforms are invertible. The inverse of R*DFT00 is R*DFT00; of R*DFT10 is R*DFT01 and vice versa (these are often called simply the DCT and IDCT, respectively); and of R*DFT11 is R*DFT11. However, the transforms computed by FFTW are unnormalized, exactly like the corresponding real and complex DFTs, so computing a transform followed by its inverse yields the original array scaled by N , where N is the logical DFT size. For REDFT00, N = 2(n 1); for RODFT00, N = 2(n + 1); otherwise, N = 2n. Note that the boundary conditions of the transform output array are given by the input boundary conditions of the inverse transform. Thus, the above transforms are all inequivalent in terms of input/output boundary conditions, even neglecting the 0.5 shift difference. FFTW is most efficient when N is a product of small factors; note that this differs from the factorization of the physical size n for REDFT00 and RODFT00! There is another oddity: n=1 REDFT00 transforms correspond to N = 0, and so are not defined (the planner will return NULL). Otherwise, any positive n is supported. For the precise mathematical definitions of these transforms as used by FFTW, see Section 4.8 [What FFTW Really Computes], page 42. (For people accustomed to the
3

There are also type V-VIII transforms, which correspond to a logical DFT of odd size N , independent of whether the physical size n is odd, but we do not support these variants.

Chapter 2: Tutorial

13

DCT/DST, FFTWs definitions have a coefficient of 2 in front of the cos/sin functions so that they correspond precisely to an even/odd DFT of size N . Some authors also include additional multiplicative factors of 2 for selected inputs and outputs; this makes the transform orthogonal, but sacrifices the direct equivalence to a symmetric DFT.)

Which type do you need?

Since the required flavor of even/odd DFT depends upon your problem, you are the best judge of this choice, but we can make a few comments on relative efficiency to help you in your selection. In particular, R*DFT01 and R*DFT10 tend to be slightly faster than R*DFT11 (especially for odd sizes), while the R*DFT00 transforms are sometimes significantly slower (especially for even sizes).4 Thus, if only the boundary conditions on the transform inputs are specified, we generally recommend R*DFT10 over R*DFT00 and R*DFT01 over R*DFT11 (unless the half-sample shift or the self-inverse property is significant for your problem). If performance is important to you and you are using only small sizes (say n < 200), e.g. for multi-dimensional transforms, then you might consider generating hard-coded transforms of those sizes and types that you are interested in (see Section 10.4 [Generating your own code], page 101). We are interested in hearing what types of symmetric transforms you find most useful.

2.5.3 The Discrete Hartley Transform

If you are planning to use the DHT because youve heard that it is faster than the DFT (FFT), stop here. The DHT is not faster than the DFT. That story is an old but enduring misconception that was debunked in 1987. The discrete Hartley transform (DHT) is an invertible linear transform closely related to the DFT. In the DFT, one multiplies each input by cos i sin (a complex exponential), whereas in the DHT each input is multiplied by simply cos + sin. Thus, the DHT transforms n real numbers to n real numbers, and has the convenient property of being its own inverse. In FFTW, a DHT (of any positive n) can be specified by an r2r kind of FFTW_DHT. Like the DFT, in FFTW the DHT is unnormalized, so computing a DHT of size n followed by another DHT of the same size will result in the original array multiplied by n. The DHT was originally proposed as a more efficient alternative to the DFT for real data, but it was subsequently shown that a specialized DFT (such as FFTWs r2hc or r2c transforms) could be just as fast. In FFTW, the DHT is actually computed by post-processing an r2hc transform, so there is ordinarily no reason to prefer it from a performance per4

R*DFT00 is sometimes slower in FFTW because we discovered that the standard algorithm for computing this by a pre/post-processed real DFTthe algorithm used in FFTPACK, Numerical Recipes, and other sources for decades nowhas serious numerical problems: it already loses several decimal places of accuracy for 16k sizes. There seem to be only two alternatives in the literature that do not suffer similarly: a recursive decomposition into smaller DCTs, which would require a large set of codelets for efficiency and generality, or sacrificing a factor of 2 in speed to use a real DFT of twice the size. We currently employ the latter technique for general n, as well as a limited form of the former method: a split-radix decomposition when n is odd (N a multiple of 4). For N containing many factors of 2, the split-radix method seems to recover most of the speed of the standard algorithm without the accuracy tradeoff.

14

FFTW 3.3.3

spective.5 However, we have heard rumors that the DHT might be the most appropriate transform in its own right for certain applications, and we would be very interested to hear from anyone who finds it useful. If FFTW_DHT is specified for multiple dimensions of a multi-dimensional transform, FFTW computes the separable product of 1d DHTs along each dimension. Unfortunately, this is not quite the same thing as a true multi-dimensional DHT; you can compute the latter, if necessary, with at most rank-1 post-processing passes [see e.g. H. Hao and R. N. Bracewell, Proc. IEEE 75, 264266 (1987)]. For the precise mathematical definition of the DHT as used by FFTW, see Section 4.8 [What FFTW Really Computes], page 42.

We provide the DHT mainly as a byproduct of some internal algorithms. FFTW computes a real input/output DFT of prime size by re-expressing it as a DHT plus post/pre-processing and then using Raders prime-DFT algorithm adapted to the DHT.

Chapter 3: Other Important Topics

15

3 Other Important Topics

3.1 SIMD alignment and fftw malloc
SIMD, which stands for Single Instruction Multiple Data, is a set of special operations supported by some processors to perform a single operation on several numbers (usually 2 or 4) simultaneously. SIMD floating-point instructions are available on several popular CPUs: SSE/SSE2/AVX on recent x86/x86-64 processors, AltiVec (single precision) on some PowerPCs (Apple G4 and higher), NEON on some ARM models, and MIPS Paired Single (currently only in FFTW 3.2.x). FFTW can be compiled to support the SIMD instructions on any of these systems. A program linking to an FFTW library compiled with SIMD support can obtain a nonnegligible speedup for most complex and r2c/c2r transforms. In order to obtain this speedup, however, the arrays of complex (or real) data passed to FFTW must be specially aligned in memory (typically 16-byte aligned), and often this alignment is more stringent than that provided by the usual malloc (etc.) allocation routines. In order to guarantee proper alignment for SIMD, therefore, in case your program is ever linked against a SIMD-using FFTW, we recommend allocating your transform data with fftw_malloc and de-allocating it with fftw_free. These have exactly the same interface and behavior as malloc/free, except that for a SIMD FFTW they ensure that the returned pointer has the necessary alignment (by calling memalign or its equivalent on your OS). You are not required to use fftw_malloc. You can allocate your data in any way that you like, from malloc to new (in C++) to a fixed-size array declaration. If the array happens not to be properly aligned, FFTW will not use the SIMD extensions. Since fftw_malloc only ever needs to be used for real and complex arrays, we provide two convenient wrapper routines fftw_alloc_real(N) and fftw_alloc_complex(N) that are equivalent to (double*)fftw_malloc(sizeof(double) * N) and (fftw_complex*)fftw_ malloc(sizeof(fftw_complex) * N), respectively (or their equivalents in other precisions).

3.2 Multi-dimensional Array Format

This section describes the format in which multi-dimensional arrays are stored in FFTW. We felt that a detailed discussion of this topic was necessary. Since several different formats are common, this topic is often a source of confusion.

3.2.1 Row-major Format

The multi-dimensional arrays passed to fftw_plan_dft etcetera are expected to be stored as a single contiguous block in row-major order (sometimes called C order). Basically, this means that as you step through adjacent memory locations, the first dimensions index varies most slowly and the last dimensions index varies most quickly. To be more explicit, let us consider an array of rank d whose dimensions are n0 n1 n2 nd1 . Now, we specify a location in the array by a sequence of d (zero-based) indices, one for each dimension: (i0 , i1 , i2 , . . . , id1 ). If the array is stored in row-major order, then this element is located at the position id1 + nd1 (id2 + nd2 (. . . + n1 i0 )).

16

FFTW 3.3.3

Note that, for the ordinary complex DFT, each element of the array must be of type fftw_ complex; i.e. a (real, imaginary) pair of (double-precision) numbers. In the advanced FFTW interface, the physical dimensions n from which the indices are computed can be different from (larger than) the logical dimensions of the transform to be computed, in order to transform a subset of a larger array. Note also that, in the advanced interface, the expression above is multiplied by a stride to get the actual array indexthis is useful in situations where each element of the multi-dimensional array is actually a data structure (or another array), and you just want to transform a single field. In the basic interface, however, the stride is 1.

3.2.2 Column-major Format

Readers from the Fortran world are used to arrays stored in column-major order (sometimes called Fortran order). This is essentially the exact opposite of row-major order in that, here, the first dimensions index varies most quickly. If you have an array stored in column-major order and wish to transform it using FFTW, it is quite easy to do. When creating the plan, simply pass the dimensions of the array to the planner in reverse order. For example, if your array is a rank three N x M x L matrix in column-major order, you should pass the dimensions of the array as if it were an L x M x N matrix (which it is, from the perspective of FFTW). This is done for you automatically by the FFTW legacy-Fortran interface (see Chapter 8 [Calling FFTW from Legacy Fortran], page 87), but you must do it manually with the modern Fortran interface (see Section 7.2 [Reversing array dimensions], page 78).

3.2.3 Fixed-size Arrays in C

A multi-dimensional array whose size is declared at compile time in C is already in rowmajor order. You dont have to do anything special to transform it. For example: { fftw_complex data[N0][N1][N2]; fftw_plan plan; ... plan = fftw_plan_dft_3d(N0, N1, N2, &data[0][0][0], &data[0][0][0], FFTW_FORWARD, FFTW_ESTIMATE); ... } This will plan a 3d in-place transform of size N0 x N1 x N2. Notice how we took the address of the zero-th element to pass to the planner (we could also have used a typecast). However, we tend to discourage users from declaring their arrays in this way, for two reasons. First, this allocates the array on the stack (automatic storage), which has a very limited size on most operating systems (declaring an array with more than a few thousand elements will often cause a crash). (You can get around this limitation on many systems by declaring the array as static and/or global, but that has its own drawbacks.) Second, it may not optimally align the array for use with a SIMD FFTW (see Section 3.1 [SIMD alignment and fftw malloc], page 15). Instead, we recommend using fftw_malloc, as described below.

Chapter 3: Other Important Topics

17

3.2.4 Dynamic Arrays in C

We recommend allocating most arrays dynamically, with fftw_malloc. This isnt too hard to do, although it is not as straightforward for multi-dimensional arrays as it is for onedimensional arrays. Creating the array is simple: using a dynamic-allocation routine like fftw_malloc, allocate an array big enough to store N fftw_complex values (for a complex DFT), where N is the product of the sizes of the array dimensions (i.e. the total number of complex values in the array). For example, here is code to allocate a 5 12 27 rank-3 array: fftw_complex *an_array; an_array = (fftw_complex*) fftw_malloc(5*12*27 * sizeof(fftw_complex)); Accessing the array elements, however, is more trickyyou cant simply use multiple applications of the [] operator like you could for fixed-size arrays. Instead, you have to explicitly compute the offset into the array using the formula given earlier for row-major arrays. For example, to reference the (i, j, k )-th element of the array allocated above, you would use the expression an_array[k + 27 * (j + 12 * i)]. This pain can be alleviated somewhat by defining appropriate macros, or, in C++, creating a class and overloading the () operator. The recent C99 standard provides a way to reinterpret the dynamic array as a variable-length multi-dimensional array amenable to [], but this feature is not yet widely supported by compilers.

3.2.5 Dynamic Arrays in CThe Wrong Way

A different method for allocating multi-dimensional arrays in C is often suggested that is incompatible with FFTW: using it will cause FFTW to die a painful death. We discuss the technique here, however, because it is so commonly known and used. This method is to create arrays of pointers of arrays of pointers of . . . etcetera. For example, the analogue in this method to the example above is: int i,j; fftw_complex ***a_bad_array; /* another way to make a 5x12x27 array */

a_bad_array = (fftw_complex ***) malloc(5 * sizeof(fftw_complex **)); for (i = 0; i < 5; ++i) { a_bad_array[i] = (fftw_complex **) malloc(12 * sizeof(fftw_complex *)); for (j = 0; j < 12; ++j) a_bad_array[i][j] = (fftw_complex *) malloc(27 * sizeof(fftw_complex)); } As you can see, this sort of array is inconvenient to allocate (and deallocate). On the other hand, it has the advantage that the (i, j, k )-th element can be referenced simply by a_bad_array[i][j][k]. If you like this technique and want to maximize convenience in accessing the array, but still want to pass the array to FFTW, you can use a hybrid method. Allocate the array as one contiguous block, but also declare an array of arrays of pointers that point to appropriate

18

FFTW 3.3.3

places in the block. That sort of trick is beyond the scope of this documentation; for more information on multi-dimensional arrays in C, see the comp.lang.c FAQ.

3.3 Words of WisdomSaving Plans

FFTW implements a method for saving plans to disk and restoring them. In fact, what FFTW does is more general than just saving and loading plans. The mechanism is called wisdom. Here, we describe this feature at a high level. See Chapter 4 [FFTW Reference], page 21, for a less casual but more complete discussion of how to use wisdom in FFTW. Plans created with the FFTW_MEASURE, FFTW_PATIENT, or FFTW_EXHAUSTIVE options produce near-optimal FFT performance, but may require a long time to compute because FFTW must measure the runtime of many possible plans and select the best one. This setup is designed for the situations where so many transforms of the same size must be computed that the start-up time is irrelevant. For short initialization times, but slower transforms, we have provided FFTW_ESTIMATE. The wisdom mechanism is a way to get the best of both worlds: you compute a good plan once, save it to disk, and later reload it as many times as necessary. The wisdom mechanism can actually save and reload many plans at once, not just one. Whenever you create a plan, the FFTW planner accumulates wisdom, which is information sufficient to reconstruct the plan. After planning, you can save this information to disk by means of the function: int fftw_export_wisdom_to_filename(const char *filename); (This function returns non-zero on success.) The next time you run the program, you can restore the wisdom with fftw_import_wisdom_ from_filename (which also returns non-zero on success), and then recreate the plan using the same flags as before. int fftw_import_wisdom_from_filename(const char *filename); Wisdom is automatically used for any size to which it is applicable, as long as the planner flags are not more patient than those with which the wisdom was created. For example, wisdom created with FFTW_MEASURE can be used if you later plan with FFTW_ESTIMATE or FFTW_MEASURE, but not with FFTW_PATIENT. The wisdom is cumulative, and is stored in a global, private data structure managed internally by FFTW. The storage space required is minimal, proportional to the logarithm of the sizes the wisdom was generated from. If memory usage is a concern, however, the wisdom can be forgotten and its associated memory freed by calling: void fftw_forget_wisdom(void); Wisdom can be exported to a file, a string, or any other medium. For details, see Section 4.7 [Wisdom], page 40.

3.4 Caveats in Using Wisdom

For in much wisdom is much grief, and he that increaseth knowledge increaseth sorrow. [Ecclesiastes 1:18]

Chapter 3: Other Important Topics

19

There are pitfalls to using wisdom, in that it can negate FFTWs ability to adapt to changing hardware and other conditions. For example, it would be perfectly possible to export wisdom from a program running on one processor and import it into a program running on another processor. Doing so, however, would mean that the second program would use plans optimized for the first processor, instead of the one it is running on. It should be safe to reuse wisdom as long as the hardware and program binaries remain unchanged. (Actually, the optimal plan may change even between runs of the same binary on identical hardware, due to differences in the virtual memory environment, etcetera. Users seriously interested in performance should worry about this problem, too.) It is likely that, if the same wisdom is used for two different program binaries, even running on the same machine, the plans may be sub-optimal because of differing code alignments. It is therefore wise to recreate wisdom every time an application is recompiled. The more the underlying hardware and software changes between the creation of wisdom and its use, the greater grows the risk of sub-optimal plans. Nevertheless, if the choice is between using FFTW_ESTIMATE or using possibly-suboptimal wisdom (created on the same machine, but for a different binary), the wisdom is likely to be better. For this reason, we provide a function to import wisdom from a standard system-wide location (/etc/fftw/wisdom on Unix): int fftw_import_system_wisdom(void); FFTW also provides a standalone program, fftw-wisdom (described by its own man page on Unix) with which users can create wisdom, e.g. for a canonical set of sizes to store in the system wisdom file. See Section 4.7.4 [Wisdom Utilities], page 41.

Chapter 4: FFTW Reference

21

4 FFTW Reference
This chapter provides a complete reference for all sequential (i.e., one-processor) FFTW functions. Parallel transforms are described in later chapters.

4.1 Data Types and Files

All programs using FFTW should include its header file: #include <fftw3.h> You must also link to the FFTW library. On Unix, this means adding -lfftw3 -lm at the end of the link command.

4.1.1 Complex numbers

The default FFTW interface uses double precision for all floating-point numbers, and defines a fftw_complex type to hold complex numbers as: typedef double fftw_complex[2]; Here, the [0] element holds the real part and the [1] element holds the imaginary part. Alternatively, if you have a C compiler (such as gcc) that supports the C99 revision of the ANSI C standard, you can use Cs new native complex type (which is binary-compatible with the typedef above). In particular, if you #include <complex.h> before <fftw3.h>, then fftw_complex is defined to be the native complex type and you can manipulate it with ordinary arithmetic (e.g. x = y * (3+4*I), where x and y are fftw_complex and I is the standard symbol for the imaginary unit); C++ has its own complex<T> template class, defined in the standard <complex> header file. Reportedly, the C++ standards committee has recently agreed to mandate that the storage format used for this type be binary-compatible with the C99 type, i.e. an array T[2] with consecutive real [0] and imaginary [1] parts. (See report http://www.open-std.org/ jtc1/sc22/WG21/docs/papers/2002/n1388.pdf WG21/N1388.) Although not part of the official standard as of this writing, the proposal stated that: This solution has been tested with all current major implementations of the standard library and shown to be working. To the extent that this is true, if you have a variable complex<double> *x, you can pass it directly to FFTW via reinterpret_cast<fftw_complex*>(x).

4.1.2 Precision
You can install single and long-double precision versions of FFTW, which replace double with float and long double, respectively (see Chapter 10 [Installation and Customization], page 97). To use these interfaces, you: Link to the single/long-double libraries; on Unix, -lfftw3f or -lfftw3l instead of (or in addition to) -lfftw3. (You can link to the different-precision libraries simultaneously.) Include the same <fftw3.h> header file. Replace all lowercase instances of fftw_ with fftwf_ or fftwl_ for single or long-double precision, respectively. (fftw_complex becomes fftwf_complex, fftw_ execute becomes fftwf_execute, etcetera.)

22

FFTW 3.3.3

Uppercase names, i.e. names beginning with FFTW_, remain the same. Replace double with float or long double for subroutine parameters. Depending upon your compiler and/or hardware, long double may not be any more precise than double (or may not be supported at all, although it is standard in C99). We also support using the nonstandard __float128 quadruple-precision type provided by recent versions of gcc on 32- and 64-bit x86 hardware (see Chapter 10 [Installation and Customization], page 97). To use this type, link with -lfftw3q -lquadmath -lm (the libquadmath library provided by gcc is needed for quadruple-precision trigonometric functions) and use fftwq_ identifiers.

4.1.3 Memory Allocation

void *fftw_malloc(size_t n); void fftw_free(void *p); These are functions that behave identically to malloc and free, except that they guarantee that the returned pointer obeys any special alignment restrictions imposed by any algorithm in FFTW (e.g. for SIMD acceleration). See Section 3.1 [SIMD alignment and fftw malloc], page 15. Data allocated by fftw_malloc must be deallocated by fftw_free and not by the ordinary free. These routines simply call through to your operating systems malloc or, if necessary, its aligned equivalent (e.g. memalign), so you normally need not worry about any significant time or space overhead. You are not required to use them to allocate your data, but we strongly recommend it. Note: in C++, just as with ordinary malloc, you must typecast the output of fftw_malloc to whatever pointer type you are allocating. We also provide the following two convenience functions to allocate real and complex arrays with n elements, which are equivalent to (double *) fftw_malloc(sizeof(double) * n) and (fftw_complex *) fftw_malloc(sizeof(fftw_complex) * n), respectively: double *fftw_alloc_real(size_t n); fftw_complex *fftw_alloc_complex(size_t n); The equivalent functions in other precisions allocate arrays of n elements in that precision. e.g. fftwf_alloc_real(n) is equivalent to (float *) fftwf_malloc(sizeof(float) * n).

4.2 Using Plans

Plans for all transform types in FFTW are stored as type fftw_plan (an opaque pointer type), and are created by one of the various planning routines described in the following sections. An fftw_plan contains all information necessary to compute the transform, including the pointers to the input and output arrays. void fftw_execute(const fftw_plan plan);

Chapter 4: FFTW Reference

23

This executes the plan, to compute the corresponding transform on the arrays for which it was planned (which must still exist). The plan is not modified, and fftw_execute can be called as many times as desired. To apply a given plan to a different array, you can use the new-array execute interface. See Section 4.6 [New-array Execute Functions], page 38. fftw_execute (and equivalents) is the only function in FFTW guaranteed to be thread-safe; see Section 5.4 [Thread safety], page 51. This function: void fftw_destroy_plan(fftw_plan plan); deallocates the plan and all its associated data. FFTWs planner saves some other persistent data, such as the accumulated wisdom and a list of algorithms available in the current configuration. If you want to deallocate all of that and reset FFTW to the pristine state it was in when you started your program, you can call: void fftw_cleanup(void); After calling fftw_cleanup, all existing plans become undefined, and you should not attempt to execute them nor to destroy them. You can however create and execute/destroy new plans, in which case FFTW starts accumulating wisdom information again. fftw_cleanup does not deallocate your plans, however. To prevent memory leaks, you must still call fftw_destroy_plan before executing fftw_cleanup. Occasionally, it may useful to know FFTWs internal cost metric that it uses to compare plans to one another; this cost is proportional to an execution time of the plan, in undocumented units, if the plan was created with the FFTW_MEASURE or other timing-based options, or alternatively is a heuristic cost function for FFTW_ESTIMATE plans. (The cost values of measured and estimated plans are not comparable, being in different units. Also, costs from different FFTW versions or the same version compiled differently may not be in the same units. Plans created from wisdom have a cost of 0 since no timing measurement is performed for them. Finally, certain problems for which only one top-level algorithm was possible may have required no measurements of the cost of the whole plan, in which case fftw_cost will also return 0.) The cost metric for a given plan is returned by: double fftw_cost(const fftw_plan plan); The following two routines are provided purely for academic purposes (that is, for entertainment). void fftw_flops(const fftw_plan plan, double *add, double *mul, double *fma); Given a plan, set add, mul, and fma to an exact count of the number of floating-point additions, multiplications, and fused multiply-add operations involved in the plans execution. The total number of floating-point operations (flops) is add + mul + 2*fma, or add + mul + fma if the hardware supports fused multiply-add instructions (although the number of FMA operations is only approximate because of compiler voodoo). (The number of operations should be an integer, but we use double to avoid overflowing int for large transforms; the arguments are of type double even for single and long-double precision versions of FFTW.)

24

FFTW 3.3.3

void fftw_fprint_plan(const fftw_plan plan, FILE *output_file); void fftw_print_plan(const fftw_plan plan); This outputs a nerd-readable representation of the plan to the given file or to stdout, respectively.

4.3 Basic Interface

Recall that the FFTW API is divided into three parts1 : the basic interface computes a single transform of contiguous data, the advanced interface computes transforms of multiple or strided arrays, and the guru interface supports the most general data layouts, multiplicities, and strides. This section describes the the basic interface, which we expect to satisfy the needs of most users.

4.3.1 Complex DFTs

fftw_plan fftw_plan_dft_1d(int n0, fftw_complex *in, fftw_complex *out, int sign, unsigned flags); fftw_plan fftw_plan_dft_2d(int n0, int n1, fftw_complex *in, fftw_complex *out, int sign, unsigned flags); fftw_plan fftw_plan_dft_3d(int n0, int n1, int n2, fftw_complex *in, fftw_complex *out, int sign, unsigned flags); fftw_plan fftw_plan_dft(int rank, const int *n, fftw_complex *in, fftw_complex *out, int sign, unsigned flags); Plan a complex input/output discrete Fourier transform (DFT) in zero or more dimensions, returning an fftw_plan (see Section 4.2 [Using Plans], page 22). Once you have created a plan for a certain transform type and parameters, then creating another plan of the same type and parameters, but for different arrays, is fast and shares constant data with the first plan (if it still exists). The planner returns NULL if the plan cannot be created. In the standard FFTW distribution, the basic interface is guaranteed to return a non-NULL plan. A plan may be NULL, however, if you are using a customized FFTW configuration supporting a restricted set of transforms.

Arguments
rank is the rank of the transform (it should be the size of the array *n), and can be any non-negative integer. (See Section 2.2 [Complex Multi-Dimensional DFTs], page 5, for the definition of rank.) The _1d, _2d, and _3d planners correspond to a rank of 1, 2, and 3, respectively. The rank may be zero, which is equivalent to a rank-1 transform of size 1, i.e. a copy of one number from input to output. n0, n1, n2, or n[0..rank-1] (as appropriate for each routine) specify the size of the transform dimensions. They can be any positive integer.
1

Gallia est omnis divisa in partes tres (Julius Caesar).

Chapter 4: FFTW Reference

25

Multi-dimensional arrays are stored in row-major order with dimensions: n0 x n1; or n0 x n1 x n2; or n[0] x n[1] x ... x n[rank-1]. See Section 3.2 [Multidimensional Array Format], page 15. FFTW is best at handling sizes of the form 2a 3b 5c 7d 11e 13f , where e + f is either 0 or 1, and the other exponents are arbitrary. Other sizes are computed by means of a slow, general-purpose algorithm (which nevertheless retains O(n log n) performance even for prime sizes). It is possible to customize FFTW for different array sizes; see Chapter 10 [Installation and Customization], page 97. Transforms whose sizes are powers of 2 are especially fast. in and out point to the input and output arrays of the transform, which may be the same (yielding an in-place transform). These arrays are overwritten during planning, unless FFTW_ESTIMATE is used in the flags. (The arrays need not be initialized, but they must be allocated.) If in == out, the transform is in-place and the input array is overwritten. If in != out, the two arrays must not overlap (but FFTW does not check for this condition). sign is the sign of the exponent in the formula that defines the Fourier transform. It can be 1 (= FFTW_FORWARD) or +1 (= FFTW_BACKWARD). flags is a bitwise OR (|) of zero or more planner flags, as defined in Section 4.3.2 [Planner Flags], page 25. FFTW computes an unnormalized transform: computing a forward followed by a backward transform (or vice versa) will result in the original data multiplied by the size of the transform (the product of the dimensions). For more information, see Section 4.8 [What FFTW Really Computes], page 42.

4.3.2 Planner Flags

All of the planner routines in FFTW accept an integer flags argument, which is a bitwise OR (|) of zero or more of the flag constants defined below. These flags control the rigor (and time) of the planning process, and can also impose (or lift) restrictions on the type of transform algorithm that is employed. Important: the planner overwrites the input array during planning unless a saved plan (see Section 4.7 [Wisdom], page 40) is available for that problem, so you should initialize your input data after creating the plan. The only exceptions to this are the FFTW_ESTIMATE and FFTW_WISDOM_ONLY flags, as mentioned below. In all cases, if wisdom is available for the given problem that was created with equal-orgreater planning rigor, then the more rigorous wisdom is used. For example, in FFTW_ ESTIMATE mode any available wisdom is used, whereas in FFTW_PATIENT mode only wisdom created in patient or exhaustive mode can be used. See Section 3.3 [Words of WisdomSaving Plans], page 18.

Planning-rigor flags
FFTW_ESTIMATE specifies that, instead of actual measurements of different algorithms, a simple heuristic is used to pick a (probably sub-optimal) plan quickly. With this flag, the input/output arrays are not overwritten during planning.

26

FFTW 3.3.3

FFTW_MEASURE tells FFTW to find an optimized plan by actually computing several FFTs and measuring their execution time. Depending on your machine, this can take some time (often a few seconds). FFTW_MEASURE is the default planning option. FFTW_PATIENT is like FFTW_MEASURE, but considers a wider range of algorithms and often produces a more optimal plan (especially for large transforms), but at the expense of several times longer planning time (especially for large transforms). FFTW_EXHAUSTIVE is like FFTW_PATIENT, but considers an even wider range of algorithms, including many that we think are unlikely to be fast, to produce the most optimal plan but with a substantially increased planning time. FFTW_WISDOM_ONLY is a special planning mode in which the plan is only created if wisdom is available for the given problem, and otherwise a NULL plan is returned. This can be combined with other flags, e.g. FFTW_WISDOM_ONLY | FFTW_PATIENT creates a plan only if wisdom is available that was created in FFTW_PATIENT or FFTW_EXHAUSTIVE mode. The FFTW_WISDOM_ONLY flag is intended for users who need to detect whether wisdom is available; for example, if wisdom is not available one may wish to allocate new arrays for planning so that user data is not overwritten.

Algorithm-restriction flags
FFTW_DESTROY_INPUT specifies that an out-of-place transform is allowed to overwrite its input array with arbitrary data; this can sometimes allow more efficient algorithms to be employed. FFTW_PRESERVE_INPUT specifies that an out-of-place transform must not change its input array. This is ordinarily the default, except for c2r and hc2r (i.e. complex-to-real) transforms for which FFTW_DESTROY_INPUT is the default. In the latter cases, passing FFTW_PRESERVE_INPUT will attempt to use algorithms that do not destroy the input, at the expense of worse performance; for multi-dimensional c2r transforms, however, no input-preserving algorithms are implemented and the planner will return NULL if one is requested. FFTW_UNALIGNED specifies that the algorithm may not impose any unusual alignment requirements on the input/output arrays (i.e. no SIMD may be used). This flag is normally not necessary, since the planner automatically detects misaligned arrays. The only use for this flag is if you want to use the new-array execute interface to execute a given plan on a different array that may not be aligned like the original. (Using fftw_malloc makes this flag unnecessary even then.)

Limiting planning time

extern void fftw_set_timelimit(double seconds); This function instructs FFTW to spend at most seconds seconds (approximately) in the planner. If seconds == FFTW_NO_TIMELIMIT (the default value, which is negative), then planning time is unbounded. Otherwise, FFTW plans with a progressively wider range of algorithms until the the given time limit is reached or the given range of algorithms is explored, returning the best available plan. For example, specifying FFTW_PATIENT first plans in FFTW_ESTIMATE mode, then in FFTW_ MEASURE mode, then finally (time permitting) in FFTW_PATIENT. If FFTW_EXHAUSTIVE is specified instead, the planner will further progress to FFTW_EXHAUSTIVE mode.

Chapter 4: FFTW Reference

27

Note that the seconds argument specifies only a rough limit; in practice, the planner may use somewhat more time if the time limit is reached when the planner is in the middle of an operation that cannot be interrupted. At the very least, the planner will complete planning in FFTW_ESTIMATE mode (which is thus equivalent to a time limit of 0).

4.3.3 Real-data DFTs

fftw_plan fftw_plan_dft_r2c_1d(int n0, double *in, fftw_complex *out, unsigned flags); fftw_plan fftw_plan_dft_r2c_2d(int n0, int n1, double *in, fftw_complex *out, unsigned flags); fftw_plan fftw_plan_dft_r2c_3d(int n0, int n1, int n2, double *in, fftw_complex *out, unsigned flags); fftw_plan fftw_plan_dft_r2c(int rank, const int *n, double *in, fftw_complex *out, unsigned flags); Plan a real-input/complex-output discrete Fourier transform (DFT) in zero or more dimensions, returning an fftw_plan (see Section 4.2 [Using Plans], page 22). Once you have created a plan for a certain transform type and parameters, then creating another plan of the same type and parameters, but for different arrays, is fast and shares constant data with the first plan (if it still exists). The planner returns NULL if the plan cannot be created. A non-NULL plan is always returned by the basic interface unless you are using a customized FFTW configuration supporting a restricted set of transforms, or if you use the FFTW_PRESERVE_INPUT flag with a multidimensional out-of-place c2r transform (see below).

Arguments
rank is the rank of the transform (it should be the size of the array *n), and can be any non-negative integer. (See Section 2.2 [Complex Multi-Dimensional DFTs], page 5, for the definition of rank.) The _1d, _2d, and _3d planners correspond to a rank of 1, 2, and 3, respectively. The rank may be zero, which is equivalent to a rank-1 transform of size 1, i.e. a copy of one real number (with zero imaginary part) from input to output. n0, n1, n2, or n[0..rank-1], (as appropriate for each routine) specify the size of the transform dimensions. They can be any positive integer. This is different in general from the physical array dimensions, which are described in Section 4.3.4 [Real-data DFT Array Format], page 28. FFTW is best at handling sizes of the form 2a 3b 5c 7d 11e 13f , where e + f is either 0 or 1, and the other exponents are arbitrary. Other sizes are computed by means of a slow, general-purpose algorithm (which nevertheless retains O(n log n) performance even for prime sizes). (It is possible to customize FFTW for different array sizes; see Chapter 10 [Installation and Customization], page 97.) Transforms

28

FFTW 3.3.3

whose sizes are powers of 2 are especially fast, and it is generally beneficial for the last dimension of an r2c/c2r transform to be even. in and out point to the input and output arrays of the transform, which may be the same (yielding an in-place transform). These arrays are overwritten during planning, unless FFTW_ESTIMATE is used in the flags. (The arrays need not be initialized, but they must be allocated.) For an in-place transform, it is important to remember that the real array will require padding, described in Section 4.3.4 [Real-data DFT Array Format], page 28. flags is a bitwise OR (|) of zero or more planner flags, as defined in Section 4.3.2 [Planner Flags], page 25. The inverse transforms, taking complex input (storing the non-redundant half of a logically Hermitian array) to real output, are given by: fftw_plan fftw_plan_dft_c2r_1d(int n0, fftw_complex *in, double *out, unsigned flags); fftw_plan fftw_plan_dft_c2r_2d(int n0, int n1, fftw_complex *in, double *out, unsigned flags); fftw_plan fftw_plan_dft_c2r_3d(int n0, int n1, int n2, fftw_complex *in, double *out, unsigned flags); fftw_plan fftw_plan_dft_c2r(int rank, const int *n, fftw_complex *in, double *out, unsigned flags); The arguments are the same as for the r2c transforms, except that the input and output data formats are reversed. FFTW computes an unnormalized transform: computing an r2c followed by a c2r transform (or vice versa) will result in the original data multiplied by the size of the transform (the product of the logical dimensions). An r2c transform produces the same output as a FFTW_ FORWARD complex DFT of the same input, and a c2r transform is correspondingly equivalent to FFTW_BACKWARD. For more information, see Section 4.8 [What FFTW Really Computes], page 42.

4.3.4 Real-data DFT Array Format

The output of a DFT of real data (r2c) contains symmetries that, in principle, make half of the outputs redundant (see Section 4.8 [What FFTW Really Computes], page 42). (Similarly for the input of an inverse c2r transform.) In practice, it is not possible to entirely realize these savings in an efficient and understandable format that generalizes to multidimensional transforms. Instead, the output of the r2c transforms is slightly over half of the output of the corresponding complex transform. We do not pack the data in any way, but store it as an ordinary array of fftw_complex values. In fact, this data is simply a subsection of what would be the array in the corresponding complex transform. Specifically, for a real transform of d (= rank) dimensions n0 n1 n2 nd1 , the complex data is an n0 n1 n2 (nd1 /2 + 1) array of fftw_complex values in

Chapter 4: FFTW Reference

29

row-major order (with the division rounded down). That is, we only store the lower half (non-negative frequencies), plus one element, of the last dimension of the data from the ordinary complex transform. (We could have instead taken half of any other dimension, but implementation turns out to be simpler if the last, contiguous, dimension is used.) For an out-of-place transform, the real data is simply an array with physical dimensions n0 n1 n2 nd1 in row-major order. For an in-place transform, some complications arise since the complex data is slightly larger than the real data. In this case, the final dimension of the real data must be padded with extra values to accommodate the size of the complex datatwo extra if the last dimension is even and one if it is odd. That is, the last dimension of the real data must physically contain 2(nd1 /2 + 1) double values (exactly enough to hold the complex data). This physical array size does not, however, change the logical array sizeonly nd1 values are actually stored in the last dimension, and nd1 is the last dimension passed to the planner.

4.3.5 Real-to-Real Transforms

fftw_plan fftw_plan_r2r_1d(int n, double *in, double *out, fftw_r2r_kind kind, unsigned flags); fftw_plan fftw_plan_r2r_2d(int n0, int n1, double *in, double *out, fftw_r2r_kind kind0, fftw_r2r_kind kind1, unsigned flags); fftw_plan fftw_plan_r2r_3d(int n0, int n1, int n2, double *in, double *out, fftw_r2r_kind kind0, fftw_r2r_kind kind1, fftw_r2r_kind kind2, unsigned flags); fftw_plan fftw_plan_r2r(int rank, const int *n, double *in, double *out, const fftw_r2r_kind *kind, unsigned flags); Plan a real input/output (r2r) transform of various kinds in zero or more dimensions, returning an fftw_plan (see Section 4.2 [Using Plans], page 22). Once you have created a plan for a certain transform type and parameters, then creating another plan of the same type and parameters, but for different arrays, is fast and shares constant data with the first plan (if it still exists). The planner returns NULL if the plan cannot be created. A non-NULL plan is always returned by the basic interface unless you are using a customized FFTW configuration supporting a restricted set of transforms, or for size-1 FFTW_REDFT00 kinds (which are not defined).

Arguments
rank is the dimensionality of the transform (it should be the size of the arrays *n and *kind), and can be any non-negative integer. The _1d, _2d, and _3d planners correspond to a rank of 1, 2, and 3, respectively. A rank of zero is equivalent to a copy of one number from input to output. n, or n0/n1/n2, or n[rank], respectively, gives the (physical) size of the transform dimensions. They can be any positive integer.

30

FFTW 3.3.3

Multi-dimensional arrays are stored in row-major order with dimensions: n0 x n1; or n0 x n1 x n2; or n[0] x n[1] x ... x n[rank-1]. See Section 3.2 [Multidimensional Array Format], page 15. FFTW is generally best at handling sizes of the form 2a 3b 5c 7d 11e 13f , where e + f is either 0 or 1, and the other exponents are arbitrary. Other sizes are computed by means of a slow, general-purpose algorithm (which nevertheless retains O(n log n) performance even for prime sizes). (It is possible to customize FFTW for different array sizes; see Chapter 10 [Installation and Customization], page 97.) Transforms whose sizes are powers of 2 are especially fast. For a REDFT00 or RODFT00 transform kind in a dimension of size n, it is n 1 or n + 1, respectively, that should be factorizable in the above form. in and out point to the input and output arrays of the transform, which may be the same (yielding an in-place transform). These arrays are overwritten during planning, unless FFTW_ESTIMATE is used in the flags. (The arrays need not be initialized, but they must be allocated.) kind, or kind0/kind1/kind2, or kind[rank], is the kind of r2r transform used for the corresponding dimension. The valid kind constants are described in Section 4.3.6 [Real-to-Real Transform Kinds], page 30. In a multi-dimensional transform, what is computed is the separable product formed by taking each transform kind along the corresponding dimension, one dimension after another. flags is a bitwise OR (|) of zero or more planner flags, as defined in Section 4.3.2 [Planner Flags], page 25.

4.3.6 Real-to-Real Transform Kinds

FFTW currently supports 11 different r2r transform kinds, specified by one of the constants below. For the precise definitions of these transforms, see Section 4.8 [What FFTW Really Computes], page 42. For a more colloquial introduction to these transform kinds, see Section 2.5 [More DFTs of Real Data], page 10. For dimension of size n, there is a corresponding logical dimension N that determines the normalization (and the optimal factorization); the formula for N is given for each kind below. Also, with each transform kind is listed its corrsponding inverse transform. FFTW computes unnormalized transforms: a transform followed by its inverse will result in the original data multiplied by N (or the product of the Ns for each dimension, in multi-dimensions). FFTW_R2HC computes a real-input DFT with output in halfcomplex format, i.e. real and imaginary parts for a transform of size n stored as: r0 , r1 , r2 , . . . , rn/2 , i(n+1)/21 , . . . , i2 , i1 (Logical N=n, inverse is FFTW_HC2R.) FFTW_HC2R computes the reverse of FFTW_R2HC, above. (Logical N=n, inverse is FFTW_ R2HC.) FFTW_DHT computes a discrete Hartley transform. (Logical N=n, inverse is FFTW_DHT.) FFTW_REDFT00 computes an REDFT00 transform, i.e. a DCT-I. (Logical N=2*(n-1), inverse is FFTW_REDFT00.)

Chapter 4: FFTW Reference

31

FFTW_REDFT10 computes an REDFT10 transform, i.e. a DCT-II (sometimes called the DCT). (Logical N=2*n, inverse is FFTW_REDFT01.) FFTW_REDFT01 computes an REDFT01 transform, i.e. a DCT-III (sometimes called the IDCT, being the inverse of DCT-II). (Logical N=2*n, inverse is FFTW_REDFT=10.) FFTW_REDFT11 computes an REDFT11 transform, i.e. a DCT-IV. (Logical N=2*n, inverse is FFTW_REDFT11.) FFTW_RODFT00 computes an RODFT00 transform, i.e. a DST-I. (Logical N=2*(n+1), inverse is FFTW_RODFT00.) FFTW_RODFT10 computes an RODFT10 transform, i.e. a DST-II. (Logical N=2*n, inverse is FFTW_RODFT01.) FFTW_RODFT01 computes an RODFT01 transform, i.e. a DST-III. (Logical N=2*n, inverse is FFTW_RODFT=10.) FFTW_RODFT11 computes an RODFT11 transform, i.e. a DST-IV. (Logical N=2*n, inverse is FFTW_RODFT11.)

4.4 Advanced Interface

FFTWs advanced interface supplements the basic interface with four new planner routines, providing a new level of flexibility: you can plan a transform of multiple arrays simultaneously, operate on non-contiguous (strided) data, and transform a subset of a larger multi-dimensional array. Other than these additional features, the planner operates in the same fashion as in the basic interface, and the resulting fftw_plan is used in the same way (see Section 4.2 [Using Plans], page 22).

4.4.1 Advanced Complex DFTs

fftw_plan fftw_plan_many_dft(int rank, const int *n, int howmany, fftw_complex *in, const int *inembed, int istride, int idist, fftw_complex *out, const int *onembed, int ostride, int odist, int sign, unsigned flags); This routine plans multiple multidimensional complex DFTs, and it extends the fftw_plan_ dft routine (see Section 4.3.1 [Complex DFTs], page 24) to compute howmany transforms, each having rank rank and size n. In addition, the transform data need not be contiguous, but it may be laid out in memory with an arbitrary stride. To account for these possibilities, fftw_plan_many_dft adds the new parameters howmany, {i,o}nembed, {i,o}stride, and {i,o}dist. The FFTW basic interface (see Section 4.3.1 [Complex DFTs], page 24) provides routines specialized for ranks 1, 2, and 3, but the advanced interface handles only the general-rank case. howmany is the number of transforms to compute. The resulting plan computes howmany transforms, where the input of the k-th transform is at location in+k*idist (in C pointer arithmetic), and its output is at location out+k*odist. Plans obtained in this way can often be faster than calling FFTW multiple times for the individual transforms. The basic

32

FFTW 3.3.3

fftw_plan_dft interface corresponds to howmany=1 (in which case the dist parameters are ignored). Each of the howmany transforms has rank rank and size n, as in the basic interface. In addition, the advanced interface allows the input and output arrays of each transform to be row-major subarrays of larger rank-rank arrays, described by inembed and onembed parameters, respectively. {i,o}nembed must be arrays of length rank, and n should be elementwise less than or equal to {i,o}nembed. Passing NULL for an nembed parameter is equivalent to passing n (i.e. same physical and logical dimensions, as in the basic interface.) The stride parameters indicate that the j-th element of the input or output arrays is located at j*istride or j*ostride, respectively. (For a multi-dimensional array, j is the ordinary row-major index.) When combined with the k-th transform in a howmany loop, from above, this means that the (j,k)-th element is at j*stride+k*dist. (The basic fftw_ plan_dft interface corresponds to a stride of 1.) For in-place transforms, the input and output stride and dist parameters should be the same; otherwise, the planner may return NULL. Arrays n, inembed, and onembed are not used after this function returns. You can safely free or reuse them. Examples: One transform of one 5 by 6 array contiguous in memory: int int int int int int rank = 2; n[] = {5, 6}; howmany = 1; idist = odist = 0; /* unused because howmany = 1 */ istride = ostride = 1; /* array is contiguous in memory */ *inembed = n, *onembed = n;

Transform of three 5 by 6 arrays, each contiguous in memory, stored in memory one after another: int int int int rank = 2; n[] = {5, 6}; howmany = 3; idist = odist = n[0]*n[1]; /* = 30, the distance in memory between the first element of the first array and the first element of the second array */ int istride = ostride = 1; /* array is contiguous in memory */ int *inembed = n, *onembed = n;

Transform each column of a 2d array with 10 rows and 3 columns: int int int int int rank = 1; /* not 2: we are computing 1d transforms */ n[] = {10}; /* 1d transforms of length 10 */ howmany = 3; idist = odist = 1; istride = ostride = 3; /* distance between two elements in the same column */ int *inembed = n, *onembed = n;

Chapter 4: FFTW Reference

33

4.4.2 Advanced Real-data DFTs

fftw_plan fftw_plan_many_dft_r2c(int rank, const int *n, int howmany, double *in, const int *inembed, int istride, int idist, fftw_complex *out, const int *onembed, int ostride, int odist, unsigned flags); fftw_plan fftw_plan_many_dft_c2r(int rank, const int *n, int howmany, fftw_complex *in, const int *inembed, int istride, int idist, double *out, const int *onembed, int ostride, int odist, unsigned flags); Like fftw_plan_many_dft, these two functions add howmany, nembed, stride, and dist parameters to the fftw_plan_dft_r2c and fftw_plan_dft_c2r functions, but otherwise behave the same as the basic interface. The interpretation of howmany, stride, and dist are the same as for fftw_plan_many_dft, above. Note that the stride and dist for the real array are in units of double, and for the complex array are in units of fftw_complex. If an nembed parameter is NULL, it is interpreted as what it would be in the basic interface, as described in Section 4.3.4 [Real-data DFT Array Format], page 28. That is, for the complex array the size is assumed to be the same as n, but with the last dimension cut roughly in half. For the real array, the size is assumed to be n if the transform is out-of-place, or n with the last dimension padded if the transform is in-place. If an nembed parameter is non-NULL, it is interpreted as the physical size of the corresponding array, in row-major order, just as for fftw_plan_many_dft. In this case, each dimension of nembed should be >= what it would be in the basic interface (e.g. the halved or padded n). Arrays n, inembed, and onembed are not used after this function returns. You can safely free or reuse them.

4.4.3 Advanced Real-to-real Transforms

fftw_plan fftw_plan_many_r2r(int rank, const int *n, int howmany, double *in, const int *inembed, int istride, int idist, double *out, const int *onembed, int ostride, int odist, const fftw_r2r_kind *kind, unsigned flags); Like fftw_plan_many_dft, this functions adds howmany, nembed, stride, and dist parameters to the fftw_plan_r2r function, but otherwise behave the same as the basic interface. The interpretation of those additional parameters are the same as for fftw_plan_many_dft. (Of course, the stride and dist parameters are now in units of double, not fftw_complex.) Arrays n, inembed, onembed, and kind are not used after this function returns. You can safely free or reuse them.

34

FFTW 3.3.3

4.5 Guru Interface

The guru interface to FFTW is intended to expose as much as possible of the flexibility in the underlying FFTW architecture. It allows one to compute multi-dimensional vectors (loops) of multi-dimensional transforms, where each vector/transform dimension has an independent size and stride. One can also use more general complex-number formats, e.g. separate real and imaginary arrays. For those users who require the flexibility of the guru interface, it is important that they pay special attention to the documentation lest they shoot themselves in the foot.

4.5.1 Interleaved and split arrays

The guru interface supports two representations of complex numbers, which we call the interleaved and the split format. The interleaved format is the same one used by the basic and advanced interfaces, and it is documented in Section 4.1.1 [Complex numbers], page 21. In the interleaved format, you provide pointers to the real part of a complex number, and the imaginary part understood to be stored in the next memory location. The split format allows separate pointers to the real and imaginary parts of a complex array. Technically, the interleaved format is redundant, because you can always express an interleaved array in terms of a split array with appropriate pointers and strides. On the other hand, the interleaved format is simpler to use, and it is common in practice. Hence, FFTW supports it as a special case.

4.5.2 Guru vector and transform sizes

The guru interface introduces one basic new data structure, fftw_iodim, that is used to specify sizes and strides for multi-dimensional transforms and vectors: typedef struct { int n; int is; int os; } fftw_iodim; Here, n is the size of the dimension, and is and os are the strides of that dimension for the input and output arrays. (The stride is the separation of consecutive elements along this dimension.) The meaning of the stride parameter depends on the type of the array that the stride refers to. If the array is interleaved complex, strides are expressed in units of complex numbers (fftw_complex). If the array is split complex or real, strides are expressed in units of real numbers (double). This convention is consistent with the usual pointer arithmetic in the C language. An interleaved array is denoted by a pointer p to fftw_complex, so that p+1 points to the next complex number. Split arrays are denoted by pointers to double, in which case pointer arithmetic operates in units of sizeof(double). The guru planner interfaces all take a (rank, dims[rank]) pair describing the transform size, and a (howmany_rank, howmany_dims[howmany_rank]) pair describing the vector

Chapter 4: FFTW Reference

35

size (a multi-dimensional loop of transforms to perform), where dims and howmany_dims are arrays of fftw_iodim. For example, the howmany parameter in the advanced complex-DFT interface corresponds to howmany_rank = 1, howmany_dims[0].n = howmany, howmany_dims[0].is = idist, and howmany_dims[0].os = odist. (To compute a single transform, you can just use howmany_rank = 0.) A row-major multidimensional array with dimensions n[rank] (see Section 3.2.1 [Rowmajor Format], page 15) corresponds to dims[i].n = n[i] and the recurrence dims[i].is = n[i+1] * dims[i+1].is (similarly for os). The stride of the last (i=rank-1) dimension is the overall stride of the array. e.g. to be equivalent to the advanced complex-DFT interface, you would have dims[rank-1].is = istride and dims[rank-1].os = ostride. In general, we only guarantee FFTW to return a non-NULL plan if the vector and transform dimensions correspond to a set of distinct indices, and for in-place transforms the input/output strides should be the same.

4.5.3 Guru Complex DFTs

fftw_plan fftw_plan_guru_dft( int rank, const fftw_iodim *dims, int howmany_rank, const fftw_iodim *howmany_dims, fftw_complex *in, fftw_complex *out, int sign, unsigned flags); fftw_plan fftw_plan_guru_split_dft( int rank, const fftw_iodim *dims, int howmany_rank, const fftw_iodim *howmany_dims, double *ri, double *ii, double *ro, double *io, unsigned flags); These two functions plan a complex-data, multi-dimensional DFT for the interleaved and split format, respectively. Transform dimensions are given by (rank, dims) over a multi-dimensional vector (loop) of dimensions (howmany_rank, howmany_dims). dims and howmany_dims should point to fftw_iodim arrays of length rank and howmany_rank, respectively. flags is a bitwise OR (|) of zero or more planner flags, as defined in Section 4.3.2 [Planner Flags], page 25. In the fftw_plan_guru_dft function, the pointers in and out point to the interleaved input and output arrays, respectively. The sign can be either 1 (= FFTW_FORWARD) or +1 (= FFTW_BACKWARD). If the pointers are equal, the transform is in-place. In the fftw_plan_guru_split_dft function, ri and ii point to the real and imaginary input arrays, and ro and io point to the real and imaginary output arrays. The input and output pointers may be the same, indicating an in-place transform. For example, for fftw_complex pointers in and out, the corresponding parameters are: ri = (double *) in; ii = (double *) in + 1;

36

FFTW 3.3.3

ro = (double *) out; io = (double *) out + 1; Because fftw_plan_guru_split_dft accepts split arrays, strides are expressed in units of double. For a contiguous fftw_complex array, the overall stride of the transform should be 2, the distance between consecutive real parts or between consecutive imaginary parts; see Section 4.5.2 [Guru vector and transform sizes], page 34. Note that the dimension strides are applied equally to the real and imaginary parts; real and imaginary arrays with different strides are not supported. There is no sign parameter in fftw_plan_guru_split_dft. This function always plans for an FFTW_FORWARD transform. To plan for an FFTW_BACKWARD transform, you can exploit the identity that the backwards DFT is equal to the forwards DFT with the real and imaginary parts swapped. For example, in the case of the fftw_complex arrays above, the FFTW_BACKWARD transform is computed by the parameters: ri ii ro io = = = = (double (double (double (double *) *) *) *) in + 1; in; out + 1; out;

4.5.4 Guru Real-data DFTs

fftw_plan fftw_plan_guru_dft_r2c( int rank, const fftw_iodim *dims, int howmany_rank, const fftw_iodim *howmany_dims, double *in, fftw_complex *out, unsigned flags); fftw_plan fftw_plan_guru_split_dft_r2c( int rank, const fftw_iodim *dims, int howmany_rank, const fftw_iodim *howmany_dims, double *in, double *ro, double *io, unsigned flags); fftw_plan fftw_plan_guru_dft_c2r( int rank, const fftw_iodim *dims, int howmany_rank, const fftw_iodim *howmany_dims, fftw_complex *in, double *out, unsigned flags); fftw_plan fftw_plan_guru_split_dft_c2r( int rank, const fftw_iodim *dims, int howmany_rank, const fftw_iodim *howmany_dims, double *ri, double *ii, double *out, unsigned flags); Plan a real-input (r2c) or real-output (c2r), multi-dimensional DFT with transform dimensions given by (rank, dims) over a multi-dimensional vector (loop) of dimensions (howmany_ rank, howmany_dims). dims and howmany_dims should point to fftw_iodim arrays of length

Chapter 4: FFTW Reference

37

rank and howmany_rank, respectively. As for the basic and advanced interfaces, an r2c transform is FFTW_FORWARD and a c2r transform is FFTW_BACKWARD. The last dimension of dims is interpreted specially: that dimension of the real array has size dims[rank-1].n, but that dimension of the complex array has size dims[rank-1].n/2+1 (division rounded down). The strides, on the other hand, are taken to be exactly as specified. It is up to the user to specify the strides appropriately for the peculiar dimensions of the data, and we do not guarantee that the planner will succeed (return non-NULL) for any dimensions other than those described in Section 4.3.4 [Real-data DFT Array Format], page 28 and generalized in Section 4.4.2 [Advanced Real-data DFTs], page 33. (That is, for an in-place transform, each individual dimension should be able to operate in place.) in and out point to the input and output arrays for r2c and c2r transforms, respectively. For split arrays, ri and ii point to the real and imaginary input arrays for a c2r transform, and ro and io point to the real and imaginary output arrays for an r2c transform. in and ro or ri and out may be the same, indicating an in-place transform. (In-place transforms where in and io or ii and out are the same are not currently supported.) flags is a bitwise OR (|) of zero or more planner flags, as defined in Section 4.3.2 [Planner Flags], page 25. In-place transforms of rank greater than 1 are currently only supported for interleaved arrays. For split arrays, the planner will return NULL.

4.5.5 Guru Real-to-real Transforms

fftw_plan fftw_plan_guru_r2r(int rank, const fftw_iodim *dims, int howmany_rank, const fftw_iodim *howmany_dims, double *in, double *out, const fftw_r2r_kind *kind, unsigned flags); Plan a real-to-real (r2r) multi-dimensional FFTW_FORWARD transform with transform dimensions given by (rank, dims) over a multi-dimensional vector (loop) of dimensions (howmany_ rank, howmany_dims). dims and howmany_dims should point to fftw_iodim arrays of length rank and howmany_rank, respectively. The transform kind of each dimension is given by the kind parameter, which should point to an array of length rank. Valid fftw_r2r_kind constants are given in Section 4.3.6 [Real-to-Real Transform Kinds], page 30. in and out point to the real input and output arrays; they may be the same, indicating an in-place transform. flags is a bitwise OR (|) of zero or more planner flags, as defined in Section 4.3.2 [Planner Flags], page 25.

4.5.6 64-bit Guru Interface

When compiled in 64-bit mode on a 64-bit architecture (where addresses are 64 bits wide), FFTW uses 64-bit quantities internally for all transform sizes, strides, and so onyou dont have to do anything special to exploit this. However, in the ordinary FFTW interfaces, you

38

FFTW 3.3.3

specify the transform size by an int quantity, which is normally only 32 bits wide. This means that, even though FFTW is using 64-bit sizes internally, you cannot specify a single transform dimension larger than 23 1 1 numbers. We expect that few users will require transforms larger than this, but, for those who do, we provide a 64-bit version of the guru interface in which all sizes are specified as integers of type ptrdiff_t instead of int. (ptrdiff_t is a signed integer type defined by the C standard to be wide enough to represent address differences, and thus must be at least 64 bits wide on a 64-bit machine.) We stress that there is no performance advantage to using this interfacethe same internal FFTW code is employed regardlessand it is only necessary if you want to specify very large transform sizes. In particular, the 64-bit guru interface is a set of planner routines that are exactly the same as the guru planner routines, except that they are named with guru64 instead of guru and they take arguments of type fftw_iodim64 instead of fftw_iodim. For example, instead of fftw_plan_guru_dft, we have fftw_plan_guru64_dft. fftw_plan fftw_plan_guru64_dft( int rank, const fftw_iodim64 *dims, int howmany_rank, const fftw_iodim64 *howmany_dims, fftw_complex *in, fftw_complex *out, int sign, unsigned flags); The fftw_iodim64 type is similar to fftw_iodim, with the same interpretation, except that it uses type ptrdiff_t instead of type int. typedef struct { ptrdiff_t n; ptrdiff_t is; ptrdiff_t os; } fftw_iodim64; Every other fftw_plan_guru function also has a fftw_plan_guru64 equivalent, but we do not repeat their documentation here since they are identical to the 32-bit versions except as noted above.

4.6 New-array Execute Functions

Normally, one executes a plan for the arrays with which the plan was created, by calling fftw_execute(plan) as described in Section 4.2 [Using Plans], page 22. However, it is possible for sophisticated users to apply a given plan to a different array using the newarray execute functions detailed below, provided that the following conditions are met: The array size, strides, etcetera are the same (since those are set by the plan). The input and output arrays are the same (in-place) or different (out-of-place) if the plan was originally created to be in-place or out-of-place, respectively. For split arrays, the separations between the real and imaginary parts, ii-ri and ioro, are the same as they were for the input and output arrays when the plan was created. (This condition is automatically satisfied for interleaved arrays.) The alignment of the new input/output arrays is the same as that of the input/output arrays when the plan was created, unless the plan was created with the FFTW_UNALIGNED

Chapter 4: FFTW Reference

39

flag. Here, the alignment is a platform-dependent quantity (for example, it is the address modulo 16 if SSE SIMD instructions are used, but the address modulo 4 for non-SIMD single-precision FFTW on the same machine). In general, only arrays allocated with fftw_malloc are guaranteed to be equally aligned (see Section 3.1 [SIMD alignment and fftw malloc], page 15). The alignment issue is especially critical, because if you dont use fftw_malloc then you may have little control over the alignment of arrays in memory. For example, neither the C++ new function nor the Fortran allocate statement provide strong enough guarantees about data alignment. If you dont use fftw_malloc, therefore, you probably have to use FFTW_UNALIGNED (which disables most SIMD support). If possible, it is probably better for you to simply create multiple plans (creating a new plan is quick once one exists for a given size), or better yet re-use the same array for your transforms. If you are tempted to use the new-array execute interface because you want to transform a known bunch of arrays of the same size, you should probably go use the advanced interface instead (see Section 4.4 [Advanced Interface], page 31)). The new-array execute functions are: void fftw_execute_dft( const fftw_plan p, fftw_complex *in, fftw_complex *out); void fftw_execute_split_dft( const fftw_plan p, double *ri, double *ii, double *ro, double *io); void fftw_execute_dft_r2c( const fftw_plan p, double *in, fftw_complex *out); void fftw_execute_split_dft_r2c( const fftw_plan p, double *in, double *ro, double *io); void fftw_execute_dft_c2r( const fftw_plan p, fftw_complex *in, double *out); void fftw_execute_split_dft_c2r( const fftw_plan p, double *ri, double *ii, double *out); void fftw_execute_r2r( const fftw_plan p, double *in, double *out); These execute the plan to compute the corresponding transform on the input/output arrays specified by the subsequent arguments. The input/output array arguments have the same

40

FFTW 3.3.3

meanings as the ones passed to the guru planner routines in the preceding sections. The plan is not modified, and these routines can be called as many times as desired, or intermixed with calls to the ordinary fftw_execute. The plan must have been created for the transform type corresponding to the execute function, e.g. it must be a complex-DFT plan for fftw_execute_dft. Any of the planner routines for that transform type, from the basic to the guru interface, could have been used to create the plan, however.

4.7 Wisdom
This section documents the FFTW mechanism for saving and restoring plans from disk. This mechanism is called wisdom.

4.7.1 Wisdom Export

int fftw_export_wisdom_to_filename(const char *filename); void fftw_export_wisdom_to_file(FILE *output_file); char *fftw_export_wisdom_to_string(void); void fftw_export_wisdom(void (*write_char)(char c, void *), void *data); These functions allow you to export all currently accumulated wisdom in a form from which it can be later imported and restored, even during a separate run of the program. (See Section 3.3 [Words of Wisdom-Saving Plans], page 18.) The current store of wisdom is not affected by calling any of these routines. fftw_export_wisdom exports the wisdom to any output medium, as specified by the callback function write_char. write_char is a putc-like function that writes the character c to some output; its second parameter is the data pointer passed to fftw_export_wisdom. For convenience, the following three wrapper routines are provided: fftw_export_wisdom_to_filename writes wisdom to a file named filename (which is created or overwritten), returning 1 on success and 0 on failure. A lower-level function, which requires you to open and close the file yourself (e.g. if you want to write wisdom to a portion of a larger file) is fftw_export_wisdom_to_file. This writes the wisdom to the current position in output_file, which should be open with write permission; upon exit, the file remains open and is positioned at the end of the wisdom data. fftw_export_wisdom_to_string returns a pointer to a NULL-terminated string holding the wisdom data. This string is dynamically allocated, and it is the responsibility of the caller to deallocate it with free when it is no longer needed. All of these routines export the wisdom in the same format, which we will not document here except to say that it is LISP-like ASCII text that is insensitive to white space.

4.7.2 Wisdom Import

int int int int fftw_import_system_wisdom(void); fftw_import_wisdom_from_filename(const char *filename); fftw_import_wisdom_from_string(const char *input_string); fftw_import_wisdom(int (*read_char)(void *), void *data);

Chapter 4: FFTW Reference

41

These functions import wisdom into a program from data stored by the fftw_export_ wisdom functions above. (See Section 3.3 [Words of Wisdom-Saving Plans], page 18.) The imported wisdom replaces any wisdom already accumulated by the running program. fftw_import_wisdom imports wisdom from any input medium, as specified by the callback function read_char. read_char is a getc-like function that returns the next character in the input; its parameter is the data pointer passed to fftw_import_wisdom. If the end of the input data is reached (which should never happen for valid data), read_char should return EOF (as defined in <stdio.h>). For convenience, the following three wrapper routines are provided: fftw_import_wisdom_from_filename reads wisdom from a file named filename. A lowerlevel function, which requires you to open and close the file yourself (e.g. if you want to read wisdom from a portion of a larger file) is fftw_import_wisdom_from_file. This reads wisdom from the current position in input_file (which should be open with read permission); upon exit, the file remains open, but the position of the read pointer is unspecified. fftw_import_wisdom_from_string reads wisdom from the NULL-terminated string input_ string. fftw_import_system_wisdom reads wisdom from an implementation-defined standard file (/etc/fftw/wisdom on Unix and GNU systems). The return value of these import routines is 1 if the wisdom was read successfully and 0 otherwise. Note that, in all of these functions, any data in the input stream past the end of the wisdom data is simply ignored.

4.7.3 Forgetting Wisdom

void fftw_forget_wisdom(void); Calling fftw_forget_wisdom causes all accumulated wisdom to be discarded and its associated memory to be freed. (New wisdom can still be gathered subsequently, however.)

4.7.4 Wisdom Utilities

FFTW includes two standalone utility programs that deal with wisdom. We merely summarize them here, since they come with their own man pages for Unix and GNU systems (with HTML versions on our web site). The first program is fftw-wisdom (or fftwf-wisdom in single precision, etcetera), which can be used to create a wisdom file containing plans for any of the transform sizes and types supported by FFTW. It is preferable to create wisdom directly from your executable (see Section 3.4 [Caveats in Using Wisdom], page 18), but this program is useful for creating global wisdom files for fftw_import_system_wisdom. The second program is fftw-wisdom-to-conf, which takes a wisdom file as input and produces a configuration routine as output. The latter is a C subroutine that you can compile and link into your program, replacing a routine of the same name in the FFTW library, that determines which parts of FFTW are callable by your program. fftw-wisdomto-conf produces a configuration routine that links to only those parts of FFTW needed by the saved plans in the wisdom, greatly reducing the size of statically linked executables (which should only attempt to create plans corresponding to those in the wisdom, however).

42

FFTW 3.3.3

4.8 What FFTW Really Computes

In this section, we provide precise mathematical definitions for the transforms that FFTW computes. These transform definitions are fairly standard, but some authors follow slightly different conventions for the normalization of the transform (the constant factor in front) and the sign of the complex exponent. We begin by presenting the one-dimensional (1d) transform definitions, and then give the straightforward extension to multi-dimensional transforms.

4.8.1 The 1d Discrete Fourier Transform (DFT)

The forward (FFTW_FORWARD) discrete Fourier transform (DFT) of a 1d complex array X of size n computes an array Y , where:
n1

Yk =
j =0

Xj e2jk

1/n

The backward (FFTW_BACKWARD) DFT computes:

n1

Yk =
j =0

Xj e2jk

1/n

FFTW computes an unnormalized transform, in that there is no coefficient in front of the summation in the DFT. In other words, applying the forward and then the backward transform will multiply the input by n. From above, an FFTW_FORWARD transform corresponds to a sign of 1 in the exponent of the DFT. Note also that we use the standard in-order output orderingthe k -th output corresponds to the frequency k/n (or k/T , where T is your total sampling period). For those who like to think in terms of positive and negative frequencies, this means that the positive frequencies are stored in the first half of the output and the negative frequencies are stored in backwards order in the second half of the output. (The frequency k/n is the same as the frequency (n k )/n.)

4.8.2 The 1d Real-data DFT

The real-input (r2c) DFT in FFTW computes the forward transform Y of the size n real array X , exactly as defined above, i.e.
n1

Yk =
j =0

Xj e2jk

1/n

This output array Y can easily be shown to possess the Hermitian symmetry Yk = Yn k , where we take Y to be periodic so that Yn = Y0 .

As a result of this symmetry, half of the output Y is redundant (being the complex conjugate of the other half), and so the 1d r2c transforms only output elements 0. . . n/2 of Y (n/2 + 1 complex numbers), where the division by 2 is rounded down. Moreover, the Hermitian symmetry implies that Y0 and, if n is even, the Yn/2 element, are purely real. So, for the R2HC r2r transform, these elements are not stored in the halfcomplex output format.

Chapter 4: FFTW Reference

43

The c2r and H2RC r2r transforms compute the backward DFT of the complex array X with Hermitian symmetry, stored in the r2c/R2HC output formats, respectively, where the backward transform is defined exactly as for the complex case:
n1

Yk =
j =0

Xj e2jk

1/n

The outputs Y of this transform can easily be seen to be purely real, and are stored as an array of real numbers. Like FFTWs complex DFT, these transforms are unnormalized. In other words, applying the real-to-complex (forward) and then the complex-to-real (backward) transform will multiply the input by n.

4.8.3 1d Real-even DFTs (DCTs)

The Real-even symmetry DFTs in FFTW are exactly equivalent to the unnormalized forward (and backward) DFTs as defined above, where the input array X of length N is purely real and is also even symmetry. In this case, the output array is likewise real and even symmetry. For the case of REDFT00, this even symmetry means that Xj = XN j , where we take X to be periodic so that XN = X0 . Because of this redundancy, only the first n real numbers are actually stored, where N = 2(n 1). The proper definition of even symmetry for REDFT10, REDFT01, and REDFT11 transforms is somewhat more intricate because of the shifts by 1/2 of the input and/or output, although the corresponding boundary conditions are given in Section 2.5.2 [Real even/odd DFTs (cosine/sine transforms)], page 11. Because of the even symmetry, however, the sine terms in the DFT all cancel and the remaining cosine terms are written explicitly below. This formulation often leads people to call such a transform a discrete cosine transform (DCT), although it is really just a special case of the DFT. In each of the definitions below, we transform a real array X of length n to a real array Y of length n:

REDFT00 (DCT-I)
An REDFT00 transform (type-I DCT) in FFTW is defined by:
n2

Yk = X0 + (1)k Xn1 + 2
j =1

Xj cos[jk/(n 1)].

Note that this transform is not defined for n = 1. For n = 2, the summation term above is dropped as you might expect.

REDFT10 (DCT-II)
An REDFT10 transform (type-II DCT, sometimes called the DCT) in FFTW is defined by:
n1

Yk = 2
j =0

Xj cos[ (j + 1/2)k/n].

44

FFTW 3.3.3

REDFT01 (DCT-III)
An REDFT01 transform (type-III DCT) in FFTW is defined by:
n1

Yk = X0 + 2
j =1

Xj cos[j (k + 1/2)/n].

In the case of n = 1, this reduces to Y0 = X0 . Up to a scale factor (see below), this is the inverse of REDFT10 (the DCT), and so the REDFT01 (DCT-III) is sometimes called the IDCT.

REDFT11 (DCT-IV)
An REDFT11 transform (type-IV DCT) in FFTW is defined by:
n1

Yk = 2
j =0

Xj cos[ (j + 1/2)(k + 1/2)/n].

Inverses and Normalization

These definitions correspond directly to the unnormalized DFTs used elsewhere in FFTW (hence the factors of 2 in front of the summations). The unnormalized inverse of REDFT00 is REDFT00, of REDFT10 is REDFT01 and vice versa, and of REDFT11 is REDFT11. Each unnormalized inverse results in the original array multiplied by N , where N is the logical DFT size. For REDFT00, N = 2(n 1) (note that n = 1 is not defined); otherwise, N = 2n. In defining the discrete cosine transform, some authors also include additional factors of 2 (or its inverse) multiplying selected inputs and/or outputs. This is a mostly cosmetic change that makes the transform orthogonal, but sacrifices the direct equivalence to a symmetric DFT.

4.8.4 1d Real-odd DFTs (DSTs)

The Real-odd symmetry DFTs in FFTW are exactly equivalent to the unnormalized forward (and backward) DFTs as defined above, where the input array X of length N is purely real and is also odd symmetry. In this case, the output is odd symmetry and purely imaginary. For the case of RODFT00, this odd symmetry means that Xj = XN j , where we take X to be periodic so that XN = X0 . Because of this redundancy, only the first n real numbers starting at j = 1 are actually stored (the j = 0 element is zero), where N = 2(n + 1). The proper definition of odd symmetry for RODFT10, RODFT01, and RODFT11 transforms is somewhat more intricate because of the shifts by 1/2 of the input and/or output, although the corresponding boundary conditions are given in Section 2.5.2 [Real even/odd DFTs (cosine/sine transforms)], page 11. Because of the odd symmetry, however, the cosine terms in the DFT all cancel and the remaining sine terms are written explicitly below. This formulation often leads people to call such a transform a discrete sine transform (DST), although it is really just a special case of the DFT. In each of the definitions below, we transform a real array X of length n to a real array Y of length n:

Chapter 4: FFTW Reference

45

RODFT00 (DST-I)
An RODFT00 transform (type-I DST) in FFTW is defined by:
n1

Yk = 2
j =0

Xj sin[ (j + 1)(k + 1)/(n + 1)].

RODFT10 (DST-II)
An RODFT10 transform (type-II DST) in FFTW is defined by:
n1

Yk = 2
j =0

Xj sin[ (j + 1/2)(k + 1)/n].

RODFT01 (DST-III)
An RODFT01 transform (type-III DST) in FFTW is defined by:
n2

Yk = (1)k Xn1 + 2
j =0

Xj sin[ (j + 1)(k + 1/2)/n].

In the case of n = 1, this reduces to Y0 = X0 .

RODFT11 (DST-IV)
An RODFT11 transform (type-IV DST) in FFTW is defined by:
n1

Yk = 2
j =0

Xj sin[ (j + 1/2)(k + 1/2)/n].

Inverses and Normalization

These definitions correspond directly to the unnormalized DFTs used elsewhere in FFTW (hence the factors of 2 in front of the summations). The unnormalized inverse of RODFT00 is RODFT00, of RODFT10 is RODFT01 and vice versa, and of RODFT11 is RODFT11. Each unnormalized inverse results in the original array multiplied by N , where N is the logical DFT size. For RODFT00, N = 2(n + 1); otherwise, N = 2n. In defining the discrete sine transform, some authors also include additional factors of 2 (or its inverse) multiplying selected inputs and/or outputs. This is a mostly cosmetic change that makes the transform orthogonal, but sacrifices the direct equivalence to an antisymmetric DFT.

4.8.5 1d Discrete Hartley Transforms (DHTs)

The discrete Hartley transform (DHT) of a 1d real array X of size n computes a real array Y of the same size, where:
n1

Yk =
j =0

Xj [cos(2jk/n) + sin(2jk/n)].

FFTW computes an unnormalized transform, in that there is no coefficient in front of the summation in the DHT. In other words, applying the transform twice (the DHT is its own inverse) will multiply the input by n.

46

FFTW 3.3.3

4.8.6 Multi-dimensional Transforms

The multi-dimensional transforms of FFTW, in general, compute simply the separable product of the given 1d transform along each dimension of the array. Since each of these transforms is unnormalized, computing the forward followed by the backward/inverse multidimensional transform will result in the original array scaled by the product of the normalization factors for each dimension (e.g. the product of the dimension sizes, for a multidimensional DFT). As an explicit example, consider the following exact mathematical definition of our multi-dimensional DFT. Let X be a d-dimensional complex array whose elements are 2 1/ns X [j1 , j2 , . . . , jd ], where 0 js < ns for all s {1, 2, . . . , d}. Let also s = e , for all s {1, 2, . . . , d}. The forward transform computes a complex array Y , whose structure is the same as that of X , defined by
n1 1 n2 1 nd 1 jd kd j1 k1 j2 k2 d . 2 X [j1 , j2 , . . . , jd ]1 jd =0

Y [k1 , k2 , . . . , kd ] =
j1 =0 j2 =0

The backward transform computes

n1 1 n2 1 nd 1

Y [k1 , k2 , . . . , kd ] =
j1 =0 j2 =0

jd =0

jd kd j1 k1 j2 k2 . 2 d X [j1 , j2 , . . . , jd ]1

Computing the forward transform followed by the backward transform will multiply the array by d s=1 nd . The definition of FFTWs multi-dimensional DFT of real data (r2c) deserves special attention. In this case, we logically compute the full multi-dimensional DFT of the input data; since the input data are purely real, the output data have the Hermitian symmetry and therefore only one non-redundant half need be stored. More specifically, for an n0 n1 n2 nd1 multi-dimensional real-input DFT, the full (logical) complex output array Y [k0 , k1 , . . . , kd1 ] has the symmetry: Y [k0 , k1 , . . . , kd1 ] = Y [n0 k0 , n1 k1 , . . . , nd1 kd1 ] (where each dimension is periodic). Because of this symmetry, we only store the kd1 = 0 nd1 /2 elements of the last dimension (division by 2 is rounded down). (We could instead have cut any other dimension in half, but the last dimension proved computationally convenient.) This results in the peculiar array format described in more detail by Section 4.3.4 [Real-data DFT Array Format], page 28. The multi-dimensional c2r transform is simply the unnormalized inverse of the r2c transform. i.e. it is the same as FFTWs complex backward multi-dimensional DFT, operating on a Hermitian input array in the peculiar format mentioned above and outputting a real array (since the DFT output is purely real). We should remind the user that the separable product of 1d transforms along each dimension, as computed by FFTW, is not always the same thing as the usual multi-dimensional

Chapter 4: FFTW Reference

47

transform. A multi-dimensional R2HC (or HC2R) transform is not identical to the multidimensional DFT, requiring some post-processing to combine the requisite real and imaginary parts, as was described in Section 2.5.1 [The Halfcomplex-format DFT], page 11. Likewise, FFTWs multidimensional FFTW_DHT r2r transform is not the same thing as the logical multi-dimensional discrete Hartley transform defined in the literature, as discussed in Section 2.5.3 [The Discrete Hartley Transform], page 13.

Chapter 5: Multi-threaded FFTW

49

5 Multi-threaded FFTW
In this chapter we document the parallel FFTW routines for shared-memory parallel hardware. These routines, which support parallel one- and multi-dimensional transforms of both real and complex data, are the easiest way to take advantage of multiple processors with FFTW. They work just like the corresponding uniprocessor transform routines, except that you have an extra initialization routine to call, and there is a routine to set the number of threads to employ. Any program that uses the uniprocessor FFTW can therefore be trivially modified to use the multi-threaded FFTW. A shared-memory machine is one in which all CPUs can directly access the same main memory, and such machines are now common due to the ubiquity of multi-core CPUs. FFTWs multi-threading support allows you to utilize these additional CPUs transparently from a single program. However, this does not necessarily translate into performance gainswhen multiple threads/CPUs are employed, there is an overhead required for synchronization that may outweigh the computatational parallelism. Therefore, you can only benefit from threads if your problem is sufficiently large.

5.1 Installation and Supported Hardware/Software

All of the FFTW threads code is located in the threads subdirectory of the FFTW package. On Unix systems, the FFTW threads libraries and header files can be automatically configured, compiled, and installed along with the uniprocessor FFTW libraries simply by including --enable-threads in the flags to the configure script (see Section 10.1 [Installation on Unix], page 97), or --enable-openmp to use OpenMP threads. The threads routines require your operating system to have some sort of shared-memory threads support. Specifically, the FFTW threads package works with POSIX threads (available on most Unix variants, from GNU/Linux to MacOS X) and Win32 threads. OpenMP threads, which are supported in many common compilers (e.g. gcc) are also supported, and may give better performance on some systems. (OpenMP threads are also useful if you are employing OpenMP in your own code, in order to minimize conflicts between threading models.) If you have a shared-memory machine that uses a different threads API, it should be a simple matter of programming to include support for it; see the file threads/threads.c for more detail. You can compile FFTW with both --enable-threads and --enable-openmp at the same time, since they install libraries with different names (fftw3_threads and fftw3_omp, as described below). However, your programs may only link to one of these two libraries at a time. Ideally, of course, you should also have multiple processors in order to get any benefit from the threaded transforms.

5.2 Usage of Multi-threaded FFTW

Here, it is assumed that the reader is already familiar with the usage of the uniprocessor FFTW routines, described elsewhere in this manual. We only describe what one has to change in order to use the multi-threaded routines.

50

FFTW 3.3.3

First, programs using the parallel complex transforms should be linked with -lfftw3_ threads -lfftw3 -lm on Unix, or -lfftw3_omp -lfftw3 -lm if you compiled with OpenMP. You will also need to link with whatever library is responsible for threads on your system (e.g. -lpthread on GNU/Linux) or include whatever compiler flag enables OpenMP (e.g. -fopenmp with gcc). Second, before calling any FFTW routines, you should call the function: int fftw_init_threads(void); This function, which need only be called once, performs any one-time initialization required to use threads on your system. It returns zero if there was some error (which should not happen under normal circumstances) and a non-zero value otherwise. Third, before creating a plan that you want to parallelize, you should call: void fftw_plan_with_nthreads(int nthreads); The nthreads argument indicates the number of threads you want FFTW to use (or actually, the maximum number). All plans subsequently created with any planner routine will use that many threads. You can call fftw_plan_with_nthreads, create some plans, call fftw_plan_with_nthreads again with a different argument, and create some more plans for a new number of threads. Plans already created before a call to fftw_plan_with_nthreads are unaffected. If you pass an nthreads argument of 1 (the default), threads are disabled for subsequent plans. With OpenMP, to configure FFTW to use all of the currently running OpenMP threads (set by omp_set_num_threads(nthreads) or by the OMP_NUM_THREADS environment variable), you can do: fftw_plan_with_nthreads(omp_get_max_threads()). (The omp_ OpenMP functions are declared via #include <omp.h>.) Given a plan, you then execute it as usual with fftw_execute(plan), and the execution will use the number of threads specified when the plan was created. When done, you destroy it as usual with fftw_destroy_plan. As described in Section 5.4 [Thread safety], page 51, plan execution is thread-safe, but plan creation and destruction are not : you should create/destroy plans only from a single thread, but can safely execute multiple plans in parallel. There is one additional routine: if you want to get rid of all memory and other resources allocated internally by FFTW, you can call: void fftw_cleanup_threads(void); which is much like the fftw_cleanup() function except that it also gets rid of threadsrelated data. You must not execute any previously created plans after calling this function. We should also mention one other restriction: if you save wisdom from a program using the multi-threaded FFTW, that wisdom cannot be used by a program using only the singlethreaded FFTW (i.e. not calling fftw_init_threads). See Section 3.3 [Words of WisdomSaving Plans], page 18.

Chapter 5: Multi-threaded FFTW

51

5.3 How Many Threads to Use?

There is a fair amount of overhead involved in synchronizing threads, so the optimal number of threads to use depends upon the size of the transform as well as on the number of processors you have. As a general rule, you dont want to use more threads than you have processors. (Using more threads will work, but there will be extra overhead with no benefit.) In fact, if the problem size is too small, you may want to use fewer threads than you have processors. You will have to experiment with your system to see what level of parallelization is best for your problem size. Typically, the problem will have to involve at least a few thousand data points before threads become beneficial. If you plan with FFTW_PATIENT, it will automatically disable threads for sizes that dont benefit from parallelization.

5.4 Thread safety

Users writing multi-threaded programs (including OpenMP) must concern themselves with the thread safety of the libraries they usethat is, whether it is safe to call routines in parallel from multiple threads. FFTW can be used in such an environment, but some care must be taken because the planner routines share data (e.g. wisdom and trigonometric tables) between calls and plans. The upshot is that the only thread-safe (re-entrant) routine in FFTW is fftw_execute (and the new-array variants thereof). All other routines (e.g. the planner) should only be called from one thread at a time. So, for example, you can wrap a semaphore lock around any calls to the planner; even more simply, you can just create all of your plans from one thread. We do not think this should be an important restriction (FFTW is designed for the situation where the only performance-sensitive code is the actual execution of the transform), and the benefits of shared data between plans are great. Note also that, since the plan is not modified by fftw_execute, it is safe to execute the same plan in parallel by multiple threads. However, since a given plan operates by default on a fixed array, you need to use one of the new-array execute functions (see Section 4.6 [New-array Execute Functions], page 38) so that different threads compute the transform of different data. (Users should note that these comments only apply to programs using shared-memory threads or OpenMP. Parallelism using MPI or forked processes involves a separate addressspace and global variables for each process, and is not susceptible to problems of this sort.) If you are configured FFTW with the --enable-debug or --enable-debug-malloc flags (see Section 10.1 [Installation on Unix], page 97), then fftw_execute is not thread-safe. These flags are not documented because they are intended only for developing and debugging FFTW, but if you must use --enable-debug then you should also specifically pass -disable-debug-malloc for fftw_execute to be thread-safe.

Chapter 6: Distributed-memory FFTW with MPI

53

6 Distributed-memory FFTW with MPI

In this chapter we document the parallel FFTW routines for parallel systems supporting the MPI message-passing interface. Unlike the shared-memory threads described in the previous chapter, MPI allows you to use distributed-memory parallelism, where each CPU has its own separate memory, and which can scale up to clusters of many thousands of processors. This capability comes at a price, however: each process only stores a portion of the data to be transformed, which means that the data structures and programming-interface are quite different from the serial or threads versions of FFTW. Distributed-memory parallelism is especially useful when you are transforming arrays so large that they do not fit into the memory of a single processor. The storage per-process required by FFTWs MPI routines is proportional to the total array size divided by the number of processes. Conversely, distributed-memory parallelism can easily pose an unacceptably high communications overhead for small problems; the threshold problem size for which parallelism becomes advantageous will depend on the precise problem you are interested in, your hardware, and your MPI implementation. A note on terminology: in MPI, you divide the data among a set of processes which each run in their own memory address space. Generally, each process runs on a different physical processor, but this is not required. A set of processes in MPI is described by an opaque data structure called a communicator, the most common of which is the predefined communicator MPI_COMM_WORLD which refers to all processes. For more information on these and other concepts common to all MPI programs, we refer the reader to the documentation at the MPI home page. We assume in this chapter that the reader is familiar with the usage of the serial (uniprocessor) FFTW, and focus only on the concepts new to the MPI interface.

6.1 FFTW MPI Installation

All of the FFTW MPI code is located in the mpi subdirectory of the FFTW package. On Unix systems, the FFTW MPI libraries and header files are automatically configured, compiled, and installed along with the uniprocessor FFTW libraries simply by including --enable-mpi in the flags to the configure script (see Section 10.1 [Installation on Unix], page 97). Any implementation of the MPI standard, version 1 or later, should work with FFTW. The configure script will attempt to automatically detect how to compile and link code using your MPI implementation. In some cases, especially if you have multiple different MPI implementations installed or have an unusual MPI software package, you may need to provide this information explicitly. Most commonly, one compiles MPI code by invoking a special compiler command, typically mpicc for C code. The configure script knows the most common names for this command, but you can specify the MPI compilation command explicitly by setting the MPICC variable, as in ./configure MPICC=mpicc .... If, instead of a special compiler command, you need to link a certain library, you can specify the link command via the MPILIBS variable, as in ./configure MPILIBS=-lmpi

54

FFTW 3.3.3

.... Note that if your MPI library is installed in a non-standard location (one the compiler does not know about by default), you may also have to specify the location of the library and header files via LDFLAGS and CPPFLAGS variables, respectively, as in ./configure LDFLAGS=-L/path/to/mpi/libs CPPFLAGS=-I/path/to/mpi/include ....

6.2 Linking and Initializing MPI FFTW

Programs using the MPI FFTW routines should be linked with -lfftw3_mpi -lfftw3 -lm on Unix in double precision, -lfftw3f_mpi -lfftw3f -lm in single precision, and so on (see Section 4.1.2 [Precision], page 21). You will also need to link with whatever library is responsible for MPI on your system; in most MPI implementations, there is a special compiler alias named mpicc to compile and link MPI code. Before calling any FFTW routines except possibly fftw_init_threads (see Section 6.11 [Combining MPI and Threads], page 66), but after calling MPI_Init, you should call the function: void fftw_mpi_init(void); If, at the end of your program, you want to get rid of all memory and other resources allocated internally by FFTW, for both the serial and MPI routines, you can call: void fftw_mpi_cleanup(void); which is much like the fftw_cleanup() function except that it also gets rid of FFTWs MPI-related data. You must not execute any previously created plans after calling this function.

6.3 2d MPI example

Before we document the FFTW MPI interface in detail, we begin with a simple example outlining how one would perform a two-dimensional N0 by N1 complex DFT. #include <fftw3-mpi.h> int main(int argc, char **argv) { const ptrdiff_t N0 = ..., N1 = ...; fftw_plan plan; fftw_complex *data; ptrdiff_t alloc_local, local_n0, local_0_start, i, j; MPI_Init(&argc, &argv); fftw_mpi_init(); /* get local data size and allocate */ alloc_local = fftw_mpi_local_size_2d(N0, N1, MPI_COMM_WORLD, &local_n0, &local_0_start); data = fftw_alloc_complex(alloc_local); /* create plan for in-place forward DFT */

Chapter 6: Distributed-memory FFTW with MPI

55

plan = fftw_mpi_plan_dft_2d(N0, N1, data, data, MPI_COMM_WORLD, FFTW_FORWARD, FFTW_ESTIMATE); /* initialize data to some function my_function(x,y) */ for (i = 0; i < local_n0; ++i) for (j = 0; j < N1; ++j) data[i*N1 + j] = my_function(local_0_start + i, j); /* compute transforms, in-place, as many times as desired */ fftw_execute(plan); fftw_destroy_plan(plan); MPI_Finalize(); } As can be seen above, the MPI interface follows the same basic style of allocate/plan/execute/destroy as the serial FFTW routines. All of the MPI-specific routines are prefixed with fftw_mpi_ instead of fftw_. There are a few important differences, however: First, we must call fftw_mpi_init() after calling MPI_Init (required in all MPI programs) and before calling any other fftw_mpi_ routine. Second, when we create the plan with fftw_mpi_plan_dft_2d, analogous to fftw_plan_ dft_2d, we pass an additional argument: the communicator, indicating which processes will participate in the transform (here MPI_COMM_WORLD, indicating all processes). Whenever you create, execute, or destroy a plan for an MPI transform, you must call the corresponding FFTW routine on all processes in the communicator for that transform. (That is, these are collective calls.) Note that the plan for the MPI transform uses the standard fftw_execute and fftw_destroy routines (on the other hand, there are MPI-specific new-array execute functions documented below). Third, all of the FFTW MPI routines take ptrdiff_t arguments instead of int as for the serial FFTW. ptrdiff_t is a standard C integer type which is (at least) 32 bits wide on a 32-bit machine and 64 bits wide on a 64-bit machine. This is to make it easy to specify very large parallel transforms on a 64-bit machine. (You can specify 64-bit transform sizes in the serial FFTW, too, but only by using the guru64 planner interface. See Section 4.5.6 [64-bit Guru Interface], page 37.) Fourth, and most importantly, you dont allocate the entire two-dimensional array on each process. Instead, you call fftw_mpi_local_size_2d to find out what portion of the array resides on each processor, and how much space to allocate. Here, the portion of the array on each process is a local_n0 by N1 slice of the total array, starting at index local_0_start. The total number of fftw_complex numbers to allocate is given by the alloc_local return value, which may be greater than local_n0 * N1 (in case some intermediate calculations require additional storage). The data distribution in FFTWs MPI interface is described in more detail by the next section. Given the portion of the array that resides on the local process, it is straightforward to initialize the data (here to a function myfunction) and otherwise manipulate it. Of course,

56

FFTW 3.3.3

at the end of the program you may want to output the data somehow, but synchronizing this output is up to you and is beyond the scope of this manual. (One good way to output a large multi-dimensional distributed array in MPI to a portable binary file is to use the free HDF5 library; see the HDF home page.)

6.4 MPI Data Distribution

The most important concept to understand in using FFTWs MPI interface is the data distribution. With a serial or multithreaded FFT, all of the inputs and outputs are stored as a single contiguous chunk of memory. With a distributed-memory FFT, the inputs and outputs are broken into disjoint blocks, one per process. In particular, FFTW uses a 1d block distribution of the data, distributed along the first dimension. For example, if you want to perform a 100 200 complex DFT, distributed over 4 processes, each process will get a 25 200 slice of the data. That is, process 0 will get rows 0 through 24, process 1 will get rows 25 through 49, process 2 will get rows 50 through 74, and process 3 will get rows 75 through 99. If you take the same array but distribute it over 3 processes, then it is not evenly divisible so the different processes will have unequal chunks. FFTWs default choice in this case is to assign 34 rows to processes 0 and 1, and 32 rows to process 2. FFTW provides several fftw_mpi_local_size routines that you can call to find out what portion of an array is stored on the current process. In most cases, you should use the default block sizes picked by FFTW, but it is also possible to specify your own block size. For example, with a 100 200 array on three processes, you can tell FFTW to use a block size of 40, which would assign 40 rows to processes 0 and 1, and 20 rows to process 2. FFTWs default is to divide the data equally among the processes if possible, and as best it can otherwise. The rows are always assigned in rank order, i.e. process 0 gets the first block of rows, then process 1, and so on. (You can change this by using MPI_Comm_split to create a new communicator with re-ordered processes.) However, you should always call the fftw_mpi_local_size routines, if possible, rather than trying to predict FFTWs distribution choices. In particular, it is critical that you allocate the storage size that is returned by fftw_mpi_local_size, which is not necessarily the size of the local slice of the array. The reason is that intermediate steps of FFTWs algorithms involve transposing the array and redistributing the data, so at these intermediate steps FFTW may require more local storage space (albeit always proportional to the total size divided by the number of processes). The fftw_mpi_local_size functions know how much storage is required for these intermediate steps and tell you the correct amount to allocate.

6.4.1 Basic and advanced distribution interfaces

As with the planner interface, the fftw_mpi_local_size distribution interface is broken into basic and advanced (_many) interfaces, where the latter allows you to specify the block size manually and also to request block sizes when computing multiple transforms simultaneously. These functions are documented more exhaustively by the FFTW MPI Reference, but we summarize the basic ideas here using a couple of two-dimensional examples. For the 100 200 complex-DFT example, above, we would find the distribution by calling the following function in the basic interface:

Chapter 6: Distributed-memory FFTW with MPI

57

ptrdiff_t fftw_mpi_local_size_2d(ptrdiff_t n0, ptrdiff_t n1, MPI_Comm comm, ptrdiff_t *local_n0, ptrdiff_t *local_0_start); Given the total size of the data to be transformed (here, n0 = 100 and n1 = 200) and an MPI communicator (comm), this function provides three numbers. First, it describes the shape of the local data: the current process should store a local_n0 by n1 slice of the overall dataset, in row-major order (n1 dimension contiguous), starting at index local_0_start. That is, if the total dataset is viewed as a n0 by n1 matrix, the current process should store the rows local_0_start to local_0_start+local_n01. Obviously, if you are running with only a single MPI process, that process will store the entire array: local_0_start will be zero and local_n0 will be n0. See Section 3.2.1 [Row-major Format], page 15. Second, the return value is the total number of data elements (e.g., complex numbers for a complex DFT) that should be allocated for the input and output arrays on the current process (ideally with fftw_malloc or an fftw_alloc function, to ensure optimal alignment). It might seem that this should always be equal to local_n0 * n1, but this is not the case. FFTWs distributed FFT algorithms require data redistributions at intermediate stages of the transform, and in some circumstances this may require slightly larger local storage. This is discussed in more detail below, under Section 6.4.2 [Load balancing], page 58. The advanced-interface local_size function for multidimensional transforms returns the same three things (local_n0, local_0_start, and the total number of elements to allocate), but takes more inputs: ptrdiff_t fftw_mpi_local_size_many(int rnk, const ptrdiff_t *n, ptrdiff_t howmany, ptrdiff_t block0, MPI_Comm comm, ptrdiff_t *local_n0, ptrdiff_t *local_0_start); The two-dimensional case above corresponds to rnk = 2 and an array n of length 2 with n[0] = n0 and n[1] = n1. This routine is for any rnk > 1; one-dimensional transforms have their own interface because they work slightly differently, as discussed below. First, the advanced interface allows you to perform multiple transforms at once, of interleaved data, as specified by the howmany parameter. (hoamany is 1 for a single transform.) Second, here you can specify your desired block size in the n0 dimension, block0. To use FFTWs default block size, pass FFTW_MPI_DEFAULT_BLOCK (0) for block0. Otherwise, on P processes, FFTW will return local_n0 equal to block0 on the first P / block0 processes (rounded down), return local_n0 equal to n0 - block0 * (P / block0) on the next process, and local_n0 equal to zero on any remaining processes. In general, we recommend using the default block size (which corresponds to n0 / P, rounded up). For example, suppose you have P = 4 processes and n0 = 21. The default will be a block size of 6, which will give local_n0 = 6 on the first three processes and local_n0 = 3 on the last process. Instead, however, you could specify block0 = 5 if you wanted, which would give local_n0 = 5 on processes 0 to 2, local_n0 = 6 on process 3. (This choice, while it may look superficially more balanced, has the same critical path as FFTWs default but requires more communications.)

58

FFTW 3.3.3

6.4.2 Load balancing

Ideally, when you parallelize a transform over some P processes, each process should end up with work that takes equal time. Otherwise, all of the processes end up waiting on whichever process is slowest. This goal is known as load balancing. In this section, we describe the circumstances under which FFTW is able to load-balance well, and in particular how you should choose your transform size in order to load balance. Load balancing is especially difficult when you are parallelizing over heterogeneous machines; for example, if one of your processors is a old 486 and another is a Pentium IV, obviously you should give the Pentium more work to do than the 486 since the latter is much slower. FFTW does not deal with this problem, howeverit assumes that your processes run on hardware of comparable speed, and that the goal is therefore to divide the problem as equally as possible. For a multi-dimensional complex DFT, FFTW can divide the problem equally among the processes if: (i) the first dimension n0 is divisible by P ; and (ii), the product of the subsequent dimensions is divisible by P . (For the advanced interface, where you can specify multiple simultaneous transforms via some vector length howmany, a factor of howmany is included in the product of the subsequent dimensions.) For a one-dimensional complex DFT, the length N of the data should be divisible by P squared to be able to divide the problem equally among the processes.

6.4.3 Transposed distributions

Internally, FFTWs MPI transform algorithms work by first computing transforms of the data local to each process, then by globally transposing the data in some fashion to redistribute the data among the processes, transforming the new data local to each process, and transposing back. For example, a two-dimensional n0 by n1 array, distributed across the n0 dimension, is transformd by: (i) transforming the n1 dimension, which are local to each process; (ii) transposing to an n1 by n0 array, distributed across the n1 dimension; (iii) transforming the n0 dimension, which is now local to each process; (iv) transposing back. However, in many applications it is acceptable to compute a multidimensional DFT whose results are produced in transposed order (e.g., n1 by n0 in two dimensions). This provides a significant performance advantage, because it means that the final transposition step can be omitted. FFTW supports this optimization, which you specify by passing the flag FFTW_MPI_TRANSPOSED_OUT to the planner routines. To compute the inverse transform of transposed output, you specify FFTW_MPI_TRANSPOSED_IN to tell it that the input is transposed. In this section, we explain how to interpret the output format of such a transform. Suppose you have are transforming multi-dimensional data with (at least two) dimensions n0 n1 n2 nd1 . As always, it is distributed along the first dimension n0 . Now, if we compute its DFT with the FFTW_MPI_TRANSPOSED_OUT flag, the resulting output data are stored with the first two dimensions transposed: n1 n0 n2 nd1 , distributed along the n1 dimension. Conversely, if we take the n1 n0 n2 nd1 data and transform it with the FFTW_MPI_TRANSPOSED_IN flag, then the format goes back to the original n0 n1 n2 nd1 array.

Chapter 6: Distributed-memory FFTW with MPI

59

There are two ways to find the portion of the transposed array that resides on the current process. First, you can simply call the appropriate local_size function, passing n1 n0 n2 nd1 (the transposed dimensions). This would mean calling the local_size function twice, once for the transposed and once for the non-transposed dimensions. Alternatively, you can call one of the local_size_transposed functions, which returns both the non-transposed and transposed data distribution from a single call. For example, for a 3d transform with transposed output (or input), you might call: ptrdiff_t fftw_mpi_local_size_3d_transposed( ptrdiff_t n0, ptrdiff_t n1, ptrdiff_t n2, MPI_Comm comm, ptrdiff_t *local_n0, ptrdiff_t *local_0_start, ptrdiff_t *local_n1, ptrdiff_t *local_1_start); Here, local_n0 and local_0_start give the size and starting index of the n0 dimension for the non-transposed data, as in the previous sections. For transposed data (e.g. the output for FFTW_MPI_TRANSPOSED_OUT), local_n1 and local_1_start give the size and starting index of the n1 dimension, which is the first dimension of the transposed data (n1 by n0 by n2). (Note that FFTW_MPI_TRANSPOSED_IN is completely equivalent to performing FFTW_MPI_ TRANSPOSED_OUT and passing the first two dimensions to the planner in reverse order, or vice versa. If you pass both the FFTW_MPI_TRANSPOSED_IN and FFTW_MPI_TRANSPOSED_OUT flags, it is equivalent to swapping the first two dimensions passed to the planner and passing neither flag.)

6.4.4 One-dimensional distributions

For one-dimensional distributed DFTs using FFTW, matters are slightly more complicated because the data distribution is more closely tied to how the algorithm works. In particular, you can no longer pass an arbitrary block size and must accept FFTWs default; also, the block sizes may be different for input and output. Also, the data distribution depends on the flags and transform direction, in order for forward and backward transforms to work correctly. ptrdiff_t fftw_mpi_local_size_1d(ptrdiff_t n0, MPI_Comm comm, int sign, unsigned flags, ptrdiff_t *local_ni, ptrdiff_t *local_i_start, ptrdiff_t *local_no, ptrdiff_t *local_o_start); This function computes the data distribution for a 1d transform of size n0 with the given transform sign and flags. Both input and output data use block distributions. The input on the current process will consist of local_ni numbers starting at index local_i_start; e.g. if only a single process is used, then local_ni will be n0 and local_i_start will be 0. Similarly for the output, with local_no numbers starting at index local_o_start. The return value of fftw_mpi_local_size_1d will be the total number of elements to allocate on the current process (which might be slightly larger than the local size due to intermediate steps in the algorithm). As mentioned above (see Section 6.4.2 [Load balancing], page 58), the data will be divided equally among the processes if n0 is divisible by the square of the number of processes. In this case, local_ni will equal local_no. Otherwise, they may be different.

60

FFTW 3.3.3

For some applications, such as convolutions, the order of the output data is irrelevant. In this case, performance can be improved by specifying that the output data be stored in an FFTW-defined scrambled format. (In particular, this is the analogue of transposed output in the multidimensional case: scrambled output saves a communications step.) If you pass FFTW_MPI_SCRAMBLED_OUT in the flags, then the output is stored in this (undocumented) scrambled order. Conversely, to perform the inverse transform of data in scrambled order, pass the FFTW_MPI_SCRAMBLED_IN flag. In MPI FFTW, only composite sizes n0 can be parallelized; we have not yet implemented a parallel algorithm for large prime sizes.

6.5 Multi-dimensional MPI DFTs of Real Data

FFTWs MPI interface also supports multi-dimensional DFTs of real data, similar to the serial r2c and c2r interfaces. (Parallel one-dimensional real-data DFTs are not currently supported; you must use a complex transform and set the imaginary parts of the inputs to zero.) The key points to understand for r2c and c2r MPI transforms (compared to the MPI complex DFTs or the serial r2c/c2r transforms), are: Just as for serial transforms, r2c/c2r DFTs transform n0 n1 n2 nd1 real data to/from n0 n1 n2 (nd1 /2 + 1) complex data: the last dimension of the complex data is cut in half (rounded down), plus one. As for the serial transforms, the sizes you pass to the plan_dft_r2c and plan_dft_c2r are the n0 n1 n2 nd1 dimensions of the real data. Although the real data is conceptually n0 n1 n2 nd1 , it is physically stored as an n0 n1 n2 [2(nd1 /2 + 1)] array, where the last dimension has been padded to make it the same size as the complex output. This is much like the in-place serial r2c/c2r interface (see Section 2.4 [Multi-Dimensional DFTs of Real Data], page 7), except that in MPI the padding is required even for out-of-place data. The extra padding numbers are ignored by FFTW (they are not like zero-padding the transform to a larger size); they are only used to determine the data layout. The data distribution in MPI for both the real and complex data is determined by the shape of the complex data. That is, you call the appropriate local size function for the n0 n1 n2 (nd1 /2 + 1) complex data, and then use the same distribution for the real data except that the last complex dimension is replaced by a (padded) real dimension of twice the length. For example suppose we are performing an out-of-place r2c transform of L M N real data [padded to L M 2(N/2 + 1) ], resulting in L M N/2 + 1 complex data. Similar to the example in Section 6.3 [2d MPI example], page 54, we might do something like: #include <fftw3-mpi.h> int main(int argc, char **argv) { const ptrdiff_t L = ..., M = ..., N = ...; fftw_plan plan;

Chapter 6: Distributed-memory FFTW with MPI

61

double *rin; fftw_complex *cout; ptrdiff_t alloc_local, local_n0, local_0_start, i, j, k; MPI_Init(&argc, &argv); fftw_mpi_init(); /* get local data size and allocate */ alloc_local = fftw_mpi_local_size_3d(L, M, N/2+1, MPI_COMM_WORLD, &local_n0, &local_0_start); rin = fftw_alloc_real(2 * alloc_local); cout = fftw_alloc_complex(alloc_local); /* create plan for out-of-place r2c DFT */ plan = fftw_mpi_plan_dft_r2c_3d(L, M, N, rin, cout, MPI_COMM_WORLD, FFTW_MEASURE); /* initialize rin to some function my_func(x,y,z) */ for (i = 0; i < local_n0; ++i) for (j = 0; j < M; ++j) for (k = 0; k < N; ++k) rin[(i*M + j) * (2*(N/2+1)) + k] = my_func(local_0_start+i, j, k); /* compute transforms as many times as desired */ fftw_execute(plan); fftw_destroy_plan(plan); MPI_Finalize(); }

Note that we allocated rin using fftw_alloc_real with an argument of 2 * alloc_local: since alloc_local is the number of complex values to allocate, the number of real values is twice as many. The rin array is then localn 0 M 2(N/2 + 1) in row-major order, so its (i,j,k) element is at the index (i*M + j) * (2*(N/2+1)) + k (see undefined [Multidimensional Array Format ], page undefined ).

As for the complex transforms, improved performance can be obtained by specifying that the output is the transpose of the input or vice versa (see Section 6.4.3 [Transposed distributions], page 58). In our L M N r2c example, including FFTW_TRANSPOSED_OUT in the flags means that the input would be a padded L M 2(N/2 + 1) real array distributed over the L dimension, while the output would be a M L N/2 + 1 complex array distributed over the M dimension. To perform the inverse c2r transform with the same data distributions, you would use the FFTW_TRANSPOSED_IN flag.

62

FFTW 3.3.3

6.6 Other multi-dimensional Real-Data MPI Transforms

FFTWs MPI interface also supports multi-dimensional r2r transforms of all kinds supported by the serial interface (e.g. discrete cosine and sine transforms, discrete Hartley transforms, etc.). Only multi-dimensional r2r transforms, not one-dimensional transforms, are currently parallelized. These are used much like the multidimensional complex DFTs discussed above, except that the data is real rather than complex, and one needs to pass an r2r transform kind (fftw_ r2r_kind) for each dimension as in the serial FFTW (see Section 2.5 [More DFTs of Real Data], page 10). For example, one might perform a two-dimensional L M that is an REDFT10 (DCT-II) in the first dimension and an RODFT10 (DST-II) in the second dimension with code like: const ptrdiff_t L = ..., M = ...; fftw_plan plan; double *data; ptrdiff_t alloc_local, local_n0, local_0_start, i, j; /* get local data size and allocate */ alloc_local = fftw_mpi_local_size_2d(L, M, MPI_COMM_WORLD, &local_n0, &local_0_start); data = fftw_alloc_real(alloc_local); /* create plan for in-place REDFT10 x RODFT10 */ plan = fftw_mpi_plan_r2r_2d(L, M, data, data, MPI_COMM_WORLD, FFTW_REDFT10, FFTW_RODFT10, FFTW_MEASURE); /* initialize data to some function my_function(x,y) */ for (i = 0; i < local_n0; ++i) for (j = 0; j < M; ++j) data[i*M + j] = my_function(local_0_start + i, j); /* compute transforms, in-place, as many times as desired */ fftw_execute(plan); fftw_destroy_plan(plan); Notice that we use the same local_size functions as we did for complex data, only now we interpret the sizes in terms of real rather than complex values, and correspondingly use fftw_alloc_real.

6.7 FFTW MPI Transposes

The FFTWs MPI Fourier transforms rely on one or more global transposition step for their communications. For example, the multidimensional transforms work by transforming along some dimensions, then transposing to make the first dimension local and transforming that, then transposing back. Because global transposition of a block-distributed matrix has many other potential uses besides FFTs, FFTWs transpose routines can be called directly, as documented in this section.

Chapter 6: Distributed-memory FFTW with MPI

63

6.7.1 Basic distributed-transpose interface

In particular, suppose that we have an n0 by n1 array in row-major order, block-distributed across the n0 dimension. To transpose this into an n1 by n0 array block-distributed across the n1 dimension, we would create a plan by calling the following function: fftw_plan fftw_mpi_plan_transpose(ptrdiff_t n0, ptrdiff_t n1, double *in, double *out, MPI_Comm comm, unsigned flags); The input and output arrays (in and out) can be the same. The transpose is actually executed by calling fftw_execute on the plan, as usual. The flags are the usual FFTW planner flags, but support two additional flags: FFTW_ MPI_TRANSPOSED_OUT and/or FFTW_MPI_TRANSPOSED_IN. What these flags indicate, for transpose plans, is that the output and/or input, respectively, are locally transposed. That is, on each process input data is normally stored as a local_n0 by n1 array in row-major order, but for an FFTW_MPI_TRANSPOSED_IN plan the input data is stored as n1 by local_n0 in row-major order. Similarly, FFTW_MPI_TRANSPOSED_OUT means that the output is n0 by local_n1 instead of local_n1 by n0. To determine the local size of the array on each process before and after the transpose, as well as the amount of storage that must be allocated, one should call fftw_mpi_local_ size_2d_transposed, just as for a 2d DFT as described in the previous section: ptrdiff_t fftw_mpi_local_size_2d_transposed (ptrdiff_t n0, ptrdiff_t n1, MPI_Comm comm, ptrdiff_t *local_n0, ptrdiff_t *local_0_start, ptrdiff_t *local_n1, ptrdiff_t *local_1_start); Again, the return value is the local storage to allocate, which in this case is the number of real (double) values rather than complex numbers as in the previous examples.

6.7.2 Advanced distributed-transpose interface

The above routines are for a transpose of a matrix of numbers (of type double), using FFTWs default block sizes. More generally, one can perform transposes of tuples of numbers, with user-specified block sizes for the input and output: fftw_plan fftw_mpi_plan_many_transpose (ptrdiff_t n0, ptrdiff_t n1, ptrdiff_t howmany, ptrdiff_t block0, ptrdiff_t block1, double *in, double *out, MPI_Comm comm, unsigned flags); In this case, one is transposing an n0 by n1 matrix of howmany-tuples (e.g. howmany = 2 for complex numbers). The input is distributed along the n0 dimension with block size block0, and the n1 by n0 output is distributed along the n1 dimension with block size block1. If FFTW_MPI_DEFAULT_BLOCK (0) is passed for a block size then FFTW uses its default block size. To get the local size of the data on each process, you should then call fftw_mpi_local_size_many_transposed.

64

FFTW 3.3.3

6.7.3 An improved replacement for MPI Alltoall

We close this section by noting that FFTWs MPI transpose routines can be thought of as a generalization for the MPI_Alltoall function (albeit only for floating-point types), and in some circumstances can function as an improved replacement. MPI_Alltoall is defined by the MPI standard as: int MPI_Alltoall(void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int recvcnt, MPI_Datatype recvtype, MPI_Comm comm); In particular, for double* arrays in and out, consider the call: MPI_Alltoall(in, howmany, MPI_DOUBLE, out, howmany MPI_DOUBLE, comm); This is completely equivalent to:

MPI_Comm_size(comm, &P); plan = fftw_mpi_plan_many_transpose(P, P, howmany, 1, 1, in, out, comm, FFTW_ESTIMATE); fftw_execute(plan); fftw_destroy_plan(plan); That is, computing a P P transpose on P processes, with a block size of 1, is just a standard all-to-all communication. However, using the FFTW routine instead of MPI_Alltoall may have certain advantages. First of all, FFTWs routine can operate in-place (in == out) whereas MPI_Alltoall can only operate out-of-place. Second, even for out-of-place plans, FFTWs routine may be faster, especially if you need to perform the all-to-all communication many times and can afford to use FFTW_MEASURE or FFTW_PATIENT. It should certainly be no slower, not including the time to create the plan, since one of the possible algorithms that FFTW uses for an out-of-place transpose is simply to call MPI_Alltoall. However, FFTW also considers several other possible algorithms that, depending on your MPI implementation and your hardware, may be faster.

6.8 FFTW MPI Wisdom

FFTWs wisdom facility (see Section 3.3 [Words of Wisdom-Saving Plans], page 18) can be used to save MPI plans as well as to save uniprocessor plans. However, for MPI there are several unavoidable complications. First, the MPI standard does not guarantee that every process can perform file I/O (at least, not using C stdio routines)in general, we may only assume that process 0 is capable of I/O.1 So, if we want to export the wisdom from a single process to a file, we must first export the wisdom to a string, then send it to process 0, then write it to a file.
1

In fact, even this assumption is not technically guaranteed by the standard, although it seems to be universal in actual MPI implementations and is widely assumed by MPI-using software. Technically, you need to query the MPI_IO attribute of MPI_COMM_WORLD with MPI_Attr_get. If this attribute is MPI_PROC_NULL, no I/O is possible. If it is MPI_ANY_SOURCE, any process can perform I/O. Otherwise, it is the rank of a process that can perform I/O ... but since it is not guaranteed to yield the same rank on all processes, you have to do an MPI_Allreduce of some kind if you want all processes to agree about which is going to do I/O. And even then, the standard only guarantees that this process can perform output, but not input. See e.g. Parallel Programming with MPI by P. S. Pacheco, section 8.1.3. Needless to say, in our experience virtually no MPI programmers worry about this.

Chapter 6: Distributed-memory FFTW with MPI

65

Second, in principle we may want to have separate wisdom for every process, since in general the processes may run on different hardware even for a single MPI program. However, in practice FFTWs MPI code is designed for the case of homogeneous hardware (see Section 6.4.2 [Load balancing], page 58), and in this case it is convenient to use the same wisdom for every process. Thus, we need a mechanism to synchronize the wisdom. To address both of these problems, FFTW provides the following two functions: void fftw_mpi_broadcast_wisdom(MPI_Comm comm); void fftw_mpi_gather_wisdom(MPI_Comm comm); Given a communicator comm, fftw_mpi_broadcast_wisdom will broadcast the wisdom from process 0 to all other processes. Conversely, fftw_mpi_gather_wisdom will collect wisdom from all processes onto process 0. (If the plans created for the same problem by different processes are not the same, fftw_mpi_gather_wisdom will arbitrarily choose one of the plans.) Both of these functions may result in suboptimal plans for different processes if the processes are running on non-identical hardware. Both of these functions are collective calls, which means that they must be executed by all processes in the communicator. So, for example, a typical code snippet to import wisdom from a file and use it on all processes would be: { int rank; fftw_mpi_init(); MPI_Comm_rank(MPI_COMM_WORLD, &rank); if (rank == 0) fftw_import_wisdom_from_filename("mywisdom"); fftw_mpi_broadcast_wisdom(MPI_COMM_WORLD); } (Note that we must call fftw_mpi_init before importing any wisdom that might contain MPI plans.) Similarly, a typical code snippet to export wisdom from all processes to a file is: { int rank; fftw_mpi_gather_wisdom(MPI_COMM_WORLD); MPI_Comm_rank(MPI_COMM_WORLD, &rank); if (rank == 0) fftw_export_wisdom_to_filename("mywisdom"); }

6.9 Avoiding MPI Deadlocks

An MPI program can deadlock if one process is waiting for a message from another process that never gets sent. To avoid deadlocks when using FFTWs MPI routines, it is important to know which functions are collective : that is, which functions must always be called in the same order from every process in a given communicator. (For example, MPI_Barrier is the canonical example of a collective function in the MPI standard.) The functions in FFTW that are always collective are: every function beginning with fftw_mpi_plan, as well as fftw_mpi_broadcast_wisdom and fftw_mpi_gather_wisdom.

66

FFTW 3.3.3

Also, the following functions from the ordinary FFTW interface are collective when they are applied to a plan created by an fftw_mpi_plan function: fftw_execute, fftw_destroy_ plan, and fftw_flops.

6.10 FFTW MPI Performance Tips

In this section, we collect a few tips on getting the best performance out of FFTWs MPI transforms. First, because of the 1d block distribution, FFTWs parallelization is currently limited by the size of the first dimension. (Multidimensional block distributions may be supported by a future version.) More generally, you should ideally arrange the dimensions so that FFTW can divide them equally among the processes. See Section 6.4.2 [Load balancing], page 58. Second, if it is not too inconvenient, you should consider working with transposed output for multidimensional plans, as this saves a considerable amount of communications. See Section 6.4.3 [Transposed distributions], page 58. Third, the fastest choices are generally either an in-place transform or an out-of-place transform with the FFTW_DESTROY_INPUT flag (which allows the input array to be used as scratch space). In-place is especially beneficial if the amount of data per process is large. Fourth, if you have multiple arrays to transform at once, rather than calling FFTWs MPI transforms several times it usually seems to be faster to interleave the data and use the advanced interface. (This groups the communications together instead of requiring separate messages for each transform.)

6.11 Combining MPI and Threads

In certain cases, it may be advantageous to combine MPI (distributed-memory) and threads (shared-memory) parallelization. FFTW supports this, with certain caveats. For example, if you have a cluster of 4-processor shared-memory nodes, you may want to use threads within the nodes and MPI between the nodes, instead of MPI for all parallelization. In particular, it is possible to seamlessly combine the MPI FFTW routines with the multithreaded FFTW routines (see Chapter 5 [Multi-threaded FFTW], page 49). However, some care must be taken in the initialization code, which should look something like this: int threads_ok; int main(int argc, char **argv) { int provided; MPI_Init_thread(&argc, &argv, MPI_THREAD_FUNNELED, &provided); threads_ok = provided >= MPI_THREAD_FUNNELED; if (threads_ok) threads_ok = fftw_init_threads(); fftw_mpi_init(); ... if (threads_ok) fftw_plan_with_nthreads(...);

Chapter 6: Distributed-memory FFTW with MPI

67

... MPI_Finalize(); } First, note that instead of calling MPI_Init, you should call MPI_Init_threads, which is the initialization routine defined by the MPI-2 standard to indicate to MPI that your program will be multithreaded. We pass MPI_THREAD_FUNNELED, which indicates that we will only call MPI routines from the main thread. (FFTW will launch additional threads internally, but the extra threads will not call MPI code.) (You may also pass MPI_THREAD_SERIALIZED or MPI_THREAD_MULTIPLE, which requests additional multithreading support from the MPI implementation, but this is not required by FFTW.) The provided parameter returns what level of threads support is actually supported by your MPI implementation; this must be at least MPI_THREAD_FUNNELED if you want to call the FFTW threads routines, so we define a global variable threads_ok to record this. You should only call fftw_init_threads or fftw_plan_with_nthreads if threads_ok is true. For more information on thread safety in MPI, see the MPI and Threads section of the MPI-2 standard. Second, we must call fftw_init_threads before fftw_mpi_init. This is critical for technical reasons having to do with how FFTW initializes its list of algorithms. Then, if you call fftw_plan_with_nthreads(N), every MPI process will launch (up to) N threads to parallelize its transforms. For example, in the hypothetical cluster of 4-processor nodes, you might wish to launch only a single MPI process per node, and then call fftw_plan_with_nthreads(4) on each process to use all processors in the nodes. This may or may not be faster than simply using as many MPI processes as you have processors, however. On the one hand, using threads within a node eliminates the need for explicit message passing within the node. On the other hand, FFTWs transpose routines are not multi-threaded, and this means that the communications that do take place will not benefit from parallelization within the node. Moreover, many MPI implementations already have optimizations to exploit shared memory when it is available, so adding the multithreaded FFTW on top of this may be superfluous.

6.12 FFTW MPI Reference

This chapter provides a complete reference to all FFTW MPI functions, datatypes, and constants. See also Chapter 4 [FFTW Reference], page 21 for information on functions and types in common with the serial interface.

6.12.1 MPI Files and Data Types

All programs using FFTWs MPI support should include its header file: #include <fftw3-mpi.h> Note that this header file includes the serial-FFTW fftw3.h header file, and also the mpi.h header file for MPI, so you need not include those files separately. You must also link to both the FFTW MPI library and to the serial FFTW library. On Unix, this means adding -lfftw3_mpi -lfftw3 -lm at the end of the link command.

68

FFTW 3.3.3

Different precisions are handled as in the serial interface: See Section 4.1.2 [Precision], page 21. That is, fftw_ functions become fftwf_ (in single precision) etcetera, and the libraries become -lfftw3f_mpi -lfftw3f -lm etcetera on Unix. Long-double precision is supported in MPI, but quad precision (fftwq_) is not due to the lack of MPI support for this type.

6.12.2 MPI Initialization

Before calling any other FFTW MPI (fftw_mpi_) function, and before importing any wisdom for MPI problems, you must call: void fftw_mpi_init(void); If FFTW threads support is used, however, fftw_mpi_init should be called after fftw_ init_threads (see Section 6.11 [Combining MPI and Threads], page 66). Calling fftw_ mpi_init additional times (before fftw_mpi_cleanup) has no effect. If you want to deallocate all persistent data and reset FFTW to the pristine state it was in when you started your program, you can call: void fftw_mpi_cleanup(void); (This calls fftw_cleanup, so you need not call the serial cleanup routine too, although it is safe to do so.) After calling fftw_mpi_cleanup, all existing plans become undefined, and you should not attempt to execute or destroy them. You must call fftw_mpi_init again after fftw_mpi_cleanup if you want to resume using the MPI FFTW routines.

6.12.3 Using MPI Plans

Once an MPI plan is created, you can execute and destroy it using fftw_execute, fftw_ destroy_plan, and the other functions in the serial interface that operate on generic plans (see Section 4.2 [Using Plans], page 22). The fftw_execute and fftw_destroy_plan functions, applied to MPI plans, are collective calls: they must be called for all processes in the communicator that was used to create the plan. You must not use the serial new-array plan-execution functions fftw_execute_dft and so on (see Section 4.6 [New-array Execute Functions], page 38) with MPI plans. Such functions are specialized to the problem type, and there are specific new-array execute functions for MPI plans: void void void void fftw_mpi_execute_dft(fftw_plan p, fftw_complex *in, fftw_complex *out); fftw_mpi_execute_dft_r2c(fftw_plan p, double *in, fftw_complex *out); fftw_mpi_execute_dft_c2r(fftw_plan p, fftw_complex *in, double *out); fftw_mpi_execute_r2r(fftw_plan p, double *in, double *out);

These functions have the same restrictions as those of the serial new-array execute functions. They are always safe to apply to the same in and out arrays that were used to create the plan. They can only be applied to new arrarys if those arrays have the same types, dimensions, in-placeness, and alignment as the original arrays, where the best way to ensure the same alignment is to use FFTWs fftw_malloc and related allocation functions for all arrays (see Section 4.1.3 [Memory Allocation], page 22). Note that distributed transposes (see Section 6.7 [FFTW MPI Transposes], page 62) use fftw_mpi_execute_r2r, since they count as rank-zero r2r plans from FFTWs perspective.

Chapter 6: Distributed-memory FFTW with MPI

69

6.12.4 MPI Data Distribution Functions

As described above (see Section 6.4 [MPI Data Distribution], page 56), in order to allocate your arrays, before creating a plan, you must first call one of the following routines to determine the required allocation size and the portion of the array locally stored on a given process. The MPI_Comm communicator passed here must be equivalent to the communicator used below for plan creation. The basic interface for multidimensional transforms consists of the functions: ptrdiff_t fftw_mpi_local_size_2d(ptrdiff_t n0, ptrdiff_t n1, MPI_Comm comm, ptrdiff_t *local_n0, ptrdiff_t *local_0_start); ptrdiff_t fftw_mpi_local_size_3d(ptrdiff_t n0, ptrdiff_t n1, ptrdiff_t n2, MPI_Comm comm, ptrdiff_t *local_n0, ptrdiff_t *local_0_start); ptrdiff_t fftw_mpi_local_size(int rnk, const ptrdiff_t *n, MPI_Comm comm, ptrdiff_t *local_n0, ptrdiff_t *local_0_start);

ptrdiff_t fftw_mpi_local_size_2d_transposed(ptrdiff_t n0, ptrdiff_t n1, MPI_Comm comm, ptrdiff_t *local_n0, ptrdiff_t *local_0_sta ptrdiff_t *local_n1, ptrdiff_t *local_1_sta ptrdiff_t fftw_mpi_local_size_3d_transposed(ptrdiff_t n0, ptrdiff_t n1, ptrdiff_t n2, MPI_Comm comm, ptrdiff_t *local_n0, ptrdiff_t *local_0_sta ptrdiff_t *local_n1, ptrdiff_t *local_1_sta ptrdiff_t fftw_mpi_local_size_transposed(int rnk, const ptrdiff_t *n, MPI_Comm comm, ptrdiff_t *local_n0, ptrdiff_t *local_0_start, ptrdiff_t *local_n1, ptrdiff_t *local_1_start) These functions return the number of elements to allocate (complex numbers for DFT/r2c/c2r plans, real numbers for r2r plans), whereas the local_n0 and local_ 0_start return the portion (local_0_start to local_0_start + local_n0 - 1) of the first dimension of an n0 n1 n2 nd1 array that is stored on the local process. See Section 6.4.1 [Basic and advanced distribution interfaces], page 56. For FFTW_MPI_TRANSPOSED_OUT plans, the _transposed variants are useful in order to also return the local portion of the first dimension in the n1 n0 n2 nd1 transposed output. See Section 6.4.3 [Transposed distributions], page 58. The advanced interface for multidimensional transforms is:

ptrdiff_t fftw_mpi_local_size_many(int rnk, const ptrdiff_t *n, ptrdiff_t howmany, ptrdiff_t block0, MPI_Comm comm, ptrdiff_t *local_n0, ptrdiff_t *local_0_start); ptrdiff_t fftw_mpi_local_size_many_transposed(int rnk, const ptrdiff_t *n, ptrdiff_t ho ptrdiff_t block0, ptrdiff_t block1, MPI_C ptrdiff_t *local_n0, ptrdiff_t *local_0_s ptrdiff_t *local_n1, ptrdiff_t *local_1_s These differ from the basic interface in only two ways. First, they allow you to specify block sizes block0 and block1 (the latter for the transposed output); you can pass FFTW_MPI_ DEFAULT_BLOCK to use FFTWs default block size as in the basic interface. Second, you

70

FFTW 3.3.3

can pass a howmany parameter, corresponding to the advanced planning interface below: this is for transforms of contiguous howmany-tuples of numbers (howmany = 1 in the basic interface). The corresponding basic and advanced routines for one-dimensional transforms (currently only complex DFTs) are: ptrdiff_t fftw_mpi_local_size_1d( ptrdiff_t n0, MPI_Comm comm, int sign, unsigned flags, ptrdiff_t *local_ni, ptrdiff_t *local_i_start, ptrdiff_t *local_no, ptrdiff_t *local_o_start); ptrdiff_t fftw_mpi_local_size_many_1d( ptrdiff_t n0, ptrdiff_t howmany, MPI_Comm comm, int sign, unsigned flags, ptrdiff_t *local_ni, ptrdiff_t *local_i_start, ptrdiff_t *local_no, ptrdiff_t *local_o_start); As above, the return value is the number of elements to allocate (complex numbers, for complex DFTs). The local_ni and local_i_start arguments return the portion (local_ i_start to local_i_start + local_ni - 1) of the 1d array that is stored on this process for the transform input, and local_no and local_o_start are the corresponding quantities for the input. The sign (FFTW_FORWARD or FFTW_BACKWARD) and flags must match the arguments passed when creating a plan. Although the inputs and outputs have different data distributions in general, it is guaranteed that the output data distribution of an FFTW_ FORWARD plan will match the input data distribution of an FFTW_BACKWARD plan and vice versa; similarly for the FFTW_MPI_SCRAMBLED_OUT and FFTW_MPI_SCRAMBLED_IN flags. See Section 6.4.4 [One-dimensional distributions], page 59.

6.12.5 MPI Plan Creation Complex-data MPI DFTs

Plans for complex-data DFTs (see Section 6.3 [2d MPI example], page 54) are created by: fftw_plan fftw_mpi_plan_dft_1d(ptrdiff_t n0, fftw_complex *in, fftw_complex *out, MPI_Comm comm, int sign, unsigned flags); fftw_plan fftw_mpi_plan_dft_2d(ptrdiff_t n0, ptrdiff_t n1, fftw_complex *in, fftw_complex *out, MPI_Comm comm, int sign, unsigned flags); fftw_plan fftw_mpi_plan_dft_3d(ptrdiff_t n0, ptrdiff_t n1, ptrdiff_t n2, fftw_complex *in, fftw_complex *out, MPI_Comm comm, int sign, unsigned flags); fftw_plan fftw_mpi_plan_dft(int rnk, const ptrdiff_t *n, fftw_complex *in, fftw_complex *out, MPI_Comm comm, int sign, unsigned flags); fftw_plan fftw_mpi_plan_many_dft(int rnk, const ptrdiff_t *n, ptrdiff_t howmany, ptrdiff_t block, ptrdiff_t tblock, fftw_complex *in, fftw_complex *out, MPI_Comm comm, int sign, unsigned flags);

Chapter 6: Distributed-memory FFTW with MPI

71

These are similar to their serial counterparts (see Section 4.3.1 [Complex DFTs], page 24) in specifying the dimensions, sign, and flags of the transform. The comm argument gives an MPI communicator that specifies the set of processes to participate in the transform; plan creation is a collective function that must be called for all processes in the communicator. The in and out pointers refer only to a portion of the overall transform data (see Section 6.4 [MPI Data Distribution], page 56) as specified by the local_size functions in the previous section. Unless flags contains FFTW_ESTIMATE, these arrays are overwritten during plan creation as for the serial interface. For multi-dimensional transforms, any dimensions > 1 are supported; for one-dimensional transforms, only composite (non-prime) n0 are currently supported (unlike the serial FFTW). Requesting an unsupported transform size will yield a NULL plan. (As in the serial interface, highly composite sizes generally yield the best performance.) The advanced-interface fftw_mpi_plan_many_dft additionally allows you to specify the block sizes for the first dimension (block) of the n0 n1 n2 nd1 input data and the first dimension (tblock) of the n1 n0 n2 nd1 transposed data (at intermediate steps of the transform, and for the output if FFTW_TRANSPOSED_OUT is specified in flags). These must be the same block sizes as were passed to the corresponding local_size function; you can pass FFTW_MPI_DEFAULT_BLOCK to use FFTWs default block size as in the basic interface. Also, the howmany parameter specifies that the transform is of contiguous howmany-tuples rather than individual complex numbers; this corresponds to the same parameter in the serial advanced interface (see Section 4.4.1 [Advanced Complex DFTs], page 31) with stride = howmany and dist = 1.

MPI flags
The flags can be any of those for the serial FFTW (see Section 4.3.2 [Planner Flags], page 25), and in addition may include one or more of the following MPI-specific flags, which improve performance at the cost of changing the output or input data formats. FFTW_MPI_SCRAMBLED_OUT, FFTW_MPI_SCRAMBLED_IN: valid for 1d transforms only, these flags indicate that the output/input of the transform are in an undocumented scrambled order. A forward FFTW_MPI_SCRAMBLED_OUT transform can be inverted by a backward FFTW_MPI_SCRAMBLED_IN (times the usual 1/N normalization). See Section 6.4.4 [One-dimensional distributions], page 59. FFTW_MPI_TRANSPOSED_OUT, FFTW_MPI_TRANSPOSED_IN: valid for multidimensional (rnk > 1) transforms only, these flags specify that the output or input of an n0 n1 n2 nd1 transform is transposed to n1 n0 n2 nd1 . See Section 6.4.3 [Transposed distributions], page 58.

Real-data MPI DFTs

Plans for real-input/output (r2c/c2r) DFTs (see Section 6.5 [Multi-dimensional MPI DFTs of Real Data], page 60) are created by: fftw_plan fftw_mpi_plan_dft_r2c_2d(ptrdiff_t n0, ptrdiff_t n1, double *in, fftw_complex *out, MPI_Comm comm, unsigned flags); fftw_plan fftw_mpi_plan_dft_r2c_2d(ptrdiff_t n0, ptrdiff_t n1, double *in, fftw_complex *out,

72

FFTW 3.3.3

fftw_plan

fftw_plan

fftw_plan

fftw_plan

fftw_plan

fftw_plan

MPI_Comm comm, unsigned flags); fftw_mpi_plan_dft_r2c_3d(ptrdiff_t n0, ptrdiff_t n1, ptrdiff_t n2, double *in, fftw_complex *out, MPI_Comm comm, unsigned flags); fftw_mpi_plan_dft_r2c(int rnk, const ptrdiff_t *n, double *in, fftw_complex *out, MPI_Comm comm, unsigned flags); fftw_mpi_plan_dft_c2r_2d(ptrdiff_t n0, ptrdiff_t n1, fftw_complex *in, double *out, MPI_Comm comm, unsigned flags); fftw_mpi_plan_dft_c2r_2d(ptrdiff_t n0, ptrdiff_t n1, fftw_complex *in, double *out, MPI_Comm comm, unsigned flags); fftw_mpi_plan_dft_c2r_3d(ptrdiff_t n0, ptrdiff_t n1, ptrdiff_t n2, fftw_complex *in, double *out, MPI_Comm comm, unsigned flags); fftw_mpi_plan_dft_c2r(int rnk, const ptrdiff_t *n, fftw_complex *in, double *out, MPI_Comm comm, unsigned flags);

Similar to the serial interface (see Section 4.3.3 [Real-data DFTs], page 27), these transform logically n0 n1 n2 nd1 real data to/from n0 n1 n2 (nd1 /2+1) complex data, representing the non-redundant half of the conjugate-symmetry output of a real-input DFT (see Section 4.8.6 [Multi-dimensional Transforms], page 46). However, the real array must be stored within a padded n0 n1 n2 [2(nd1 /2 + 1)] array (much like the in-place serial r2c transforms, but here for out-of-place transforms as well). Currently, only multi-dimensional (rnk > 1) r2c/c2r transforms are supported (requesting a plan for rnk = 1 will yield NULL). As explained above (see Section 6.5 [Multi-dimensional MPI DFTs of Real Data], page 60), the data distribution of both the real and complex arrays is given by the local_size function called for the dimensions of the complex array. Similar to the other planning functions, the input and output arrays are overwritten when the plan is created except in FFTW_ESTIMATE mode. As for the complex DFTs above, there is an advance interface that allows you to manually specify block sizes and to transform contiguous howmany-tuples of real/complex numbers: fftw_plan fftw_mpi_plan_many_dft_r2c (int rnk, const ptrdiff_t *n, ptrdiff_t howmany, ptrdiff_t iblock, ptrdiff_t oblock, double *in, fftw_complex *out, MPI_Comm comm, unsigned flags); fftw_plan fftw_mpi_plan_many_dft_c2r (int rnk, const ptrdiff_t *n, ptrdiff_t howmany, ptrdiff_t iblock, ptrdiff_t oblock, fftw_complex *in, double *out, MPI_Comm comm, unsigned flags);

Chapter 6: Distributed-memory FFTW with MPI

73

MPI r2r transforms

There are corresponding plan-creation routines for r2r transforms (see Section 2.5 [More DFTs of Real Data], page 10), currently supporting multidimensional (rnk > 1) transforms only (rnk = 1 will yield a NULL plan): fftw_plan fftw_mpi_plan_r2r_2d(ptrdiff_t n0, ptrdiff_t n1, double *in, double *out, MPI_Comm comm, fftw_r2r_kind kind0, fftw_r2r_kind kind1, unsigned flags); fftw_plan fftw_mpi_plan_r2r_3d(ptrdiff_t n0, ptrdiff_t n1, ptrdiff_t n2, double *in, double *out, MPI_Comm comm, fftw_r2r_kind kind0, fftw_r2r_kind kind1, fftw_r2r_kind unsigned flags); fftw_plan fftw_mpi_plan_r2r(int rnk, const ptrdiff_t *n, double *in, double *out, MPI_Comm comm, const fftw_r2r_kind *kind, unsigned flags); fftw_plan fftw_mpi_plan_many_r2r(int rnk, const ptrdiff_t *n, ptrdiff_t iblock, ptrdiff_t oblock, double *in, double *out, MPI_Comm comm, const fftw_r2r_kind *kind, unsigned flags); The parameters are much the same as for the complex DFTs above, except that the arrays are of real numbers (and hence the outputs of the local_size data-distribution functions should be interpreted as counts of real rather than complex numbers). Also, the kind parameters specify the r2r kinds along each dimension as for the serial interface (see Section 4.3.6 [Real-to-Real Transform Kinds], page 30). See Section 6.6 [Other Multi-dimensional Real-data MPI Transforms], page 62.

MPI transposition
FFTW also provides routines to plan a transpose of a distributed n0 by n1 array of real numbers, or an array of howmany-tuples of real numbers with specified block sizes (see Section 6.7 [FFTW MPI Transposes], page 62): fftw_plan fftw_mpi_plan_transpose(ptrdiff_t n0, ptrdiff_t n1, double *in, double *out, MPI_Comm comm, unsigned flags); fftw_plan fftw_mpi_plan_many_transpose (ptrdiff_t n0, ptrdiff_t n1, ptrdiff_t howmany, ptrdiff_t block0, ptrdiff_t block1, double *in, double *out, MPI_Comm comm, unsigned flags); These plans are used with the fftw_mpi_execute_r2r new-array execute function (see undefined [Using MPI Plans ], page undefined ), since they count as (rank zero) r2r plans from FFTWs perspective.

74

FFTW 3.3.3

6.12.6 MPI Wisdom Communication

To facilitate synchronizing wisdom among the different MPI processes, we provide two functions: void fftw_mpi_gather_wisdom(MPI_Comm comm); void fftw_mpi_broadcast_wisdom(MPI_Comm comm); The fftw_mpi_gather_wisdom function gathers all wisdom in the given communicator comm to the process of rank 0 in the communicator: that process obtains the union of all wisdom on all the processes. As a side effect, some other processes will gain additional wisdom from other processes, but only process 0 will gain the complete union. The fftw_mpi_broadcast_wisdom does the reverse: it exports wisdom from process 0 in comm to all other processes in the communicator, replacing any wisdom they currently have. See Section 6.8 [FFTW MPI Wisdom], page 64.

6.13 FFTW MPI Fortran Interface

The FFTW MPI interface is callable from modern Fortran compilers supporting the Fortran 2003 iso_c_binding standard for calling C functions. As described in Chapter 7 [Calling FFTW from Modern Fortran], page 77, this means that you can directly call FFTWs C interface from Fortran with only minor changes in syntax. There are, however, a few things specific to the MPI interface to keep in mind: Instead of including fftw3.f03 as in undefined [Overview of Fortran interface ], page undefined , you should include fftw3-mpi.f03 (after use, intrinsic :: iso_c_binding as before). The fftw3-mpi.f03 file includes fftw3.f03, so you should not include them both yourself. (You will also want to include the MPI header file, usually via include mpif.h or similar, although though this is not needed by fftw3mpi.f03 per se.) (To use the fftwl_ long double extended-precision routines in supporting compilers, you should include fftw3f-mpi.f03 in addition to fftw3-mpi.f03. See Section 7.1.1 [Extended and quadruple precision in Fortran], page 78.) Because of the different storage conventions between C and Fortran, you reverse the order of your array dimensions when passing them to FFTW (see Section 7.2 [Reversing array dimensions], page 78). This is merely a difference in notation and incurs no performance overhead. However, it means that, whereas in C the first dimension is distributed, in Fortran the last dimension of your array is distributed. In Fortran, communicators are stored as integer types; there is no MPI_Comm type, nor is there any way to access a C MPI_Comm. Fortunately, this is taken care of for you by the FFTW Fortran interface: whenever the C interface expects an MPI_Comm type, you should pass the Fortran communicator as an integer.2 Because you need to call the local_size function to find out how much space to allocate, and this may be larger than the local portion of the array (see Section 6.4 [MPI Data Distribution], page 56), you should always allocate your arrays dynamically
2

Technically, this is because you arent actually calling the C functions directly. You are calling wrapper functions that translate the communicator with MPI_Comm_f2c before calling the ordinary C interface. This is all done transparently, however, since the fftw3-mpi.f03 interface file renames the wrappers so that they are called in Fortran with the same names as the C interface functions.

Chapter 6: Distributed-memory FFTW with MPI

75

using FFTWs allocation routines as described in Section 7.5 [Allocating aligned memory in Fortran], page 82. (Coincidentally, this also provides the best performance by guaranteeding proper data alignment.) Because all sizes in the MPI FFTW interface are declared as ptrdiff_t in C, you should use integer(C_INTPTR_T) in Fortran (see Section 7.3 [FFTW Fortran type reference], page 80). In Fortran, because of the language semantics, we generally recommend using the new-array execute functions for all plans, even in the common case where you are executing the plan on the same arrays for which the plan was created (see Section 7.4 [Plan execution in Fortran], page 81). However, note that in the MPI interface these functions are changed: fftw_execute_dft becomes fftw_mpi_execute_dft, etcetera. See Section 6.12.3 [Using MPI Plans], page 68. For example, here is a Fortran code snippet to perform a distributed L M complex DFT in-place. (This assumes you have already initialized MPI with MPI_init and have also performed call fftw_mpi_init.) use, intrinsic :: iso_c_binding include fftw3-mpi.f03 integer(C_INTPTR_T), parameter :: L = ... integer(C_INTPTR_T), parameter :: M = ... type(C_PTR) :: plan, cdata complex(C_DOUBLE_COMPLEX), pointer :: data(:,:) integer(C_INTPTR_T) :: i, j, alloc_local, local_M, local_j_offset ! get local data size and allocate (note dimension reversal) alloc_local = fftw_mpi_local_size_2d(M, L, MPI_COMM_WORLD, & local_M, local_j_offset) cdata = fftw_alloc_complex(alloc_local) call c_f_pointer(cdata, data, [L,local_M]) create MPI plan for in-place forward DFT (note dimension reversal) plan = fftw_mpi_plan_dft_2d(M, L, data, data, MPI_COMM_WORLD, & FFTW_FORWARD, FFTW_MEASURE)

! initialize data to some function my_function(i,j) do j = 1, local_M do i = 1, L data(i, j) = my_function(i, j + local_j_offset) end do end do ! compute transform (as many times as desired) call fftw_mpi_execute_dft(plan, data, data) call fftw_destroy_plan(plan) call fftw_free(cdata)

76

FFTW 3.3.3

Note that when we called fftw_mpi_local_size_2d and fftw_mpi_plan_dft_2d with the dimensions in reversed order, since a L M Fortran array is viewed by FFTW in C as a M L array. This means that the array was distributed over the M dimension, the local portion of which is a L localM array in Fortran. (You must not use an allocate statement to allocate an L localM array, however; you must allocate alloc_local complex numbers, which may be greater than L * local_M, in order to reserve space for intermediate steps of the transform.) Finally, we mention that because Cs array indices are zero-based, the local_j_offset argument can conveniently be interpreted as an offset in the 1-based j index (rather than as a starting index as in C). If instead you had used the ior(FFTW_MEASURE, FFTW_MPI_TRANSPOSED_OUT) flag, the output of the transform would be a transposed M localL array, associated with the same cdata allocation (since the transform is in-place), and which you could declare with: complex(C_DOUBLE_COMPLEX), pointer :: tdata(:,:) ... call c_f_pointer(cdata, tdata, [M,local_L]) where local_L would have been obtained by changing the fftw_mpi_local_size_2d call to: alloc_local = fftw_mpi_local_size_2d_transposed(M, L, MPI_COMM_WORLD, & local_M, local_j_offset, local_L, local_i_offset)

Chapter 7: Calling FFTW from Modern Fortran

77

7 Calling FFTW from Modern Fortran

Fortran 2003 standardized ways for Fortran code to call C libraries, and this allows us to support a direct translation of the FFTW C API into Fortran. Compared to the legacy Fortran 77 interface (see Chapter 8 [Calling FFTW from Legacy Fortran], page 87), this direct interface offers many advantages, especially compile-time type-checking and aligned memory allocation. As of this writing, support for these C interoperability features seems widespread, having been implemented in nearly all major Fortran compilers (e.g. GNU, Intel, IBM, Oracle/Solaris, Portland Group, NAG). This chapter documents that interface. For the most part, since this interface allows Fortran to call the C interface directly, the usage is identical to C translated to Fortran syntax. However, there are a few subtle points such as memory allocation, wisdom, and data types that deserve closer attention.

7.1 Overview of Fortran interface

FFTW provides a file fftw3.f03 that defines Fortran 2003 interfaces for all of its C routines, except for the MPI routines described elsewhere, which can be found in the same directory as fftw3.h (the C header file). In any Fortran subroutine where you want to use FFTW functions, you should begin with: use, intrinsic :: iso_c_binding include fftw3.f03 This includes the interface definitions and the standard iso_c_binding module (which defines the equivalents of C types). You can also put the FFTW functions into a module if you prefer (see Section 7.7 [Defining an FFTW module], page 85). At this point, you can now call anything in the FFTW C interface directly, almost exactly as in C other than minor changes in syntax. For example: type(C_PTR) :: plan complex(C_DOUBLE_COMPLEX), dimension(1024,1000) :: in, out plan = fftw_plan_dft_2d(1000,1024, in,out, FFTW_FORWARD,FFTW_ESTIMATE) ... call fftw_execute_dft(plan, in, out) ... call fftw_destroy_plan(plan) A few important things to keep in mind are: FFTW plans are type(C_PTR). Other C types are mapped in the obvious way via the iso_c_binding standard: int turns into integer(C_INT), fftw_complex turns into complex(C_DOUBLE_COMPLEX), double turns into real(C_DOUBLE), and so on. See Section 7.3 [FFTW Fortran type reference], page 80. Functions in C become functions in Fortran if they have a return value, and subroutines in Fortran otherwise. The ordering of the Fortran array dimensions must be reversed when they are passed to the FFTW plan creation, thanks to differences in array indexing conventions (see

78

FFTW 3.3.3

Section 3.2 [Multi-dimensional Array Format], page 15). This is unlike the legacy Fortran interface (see Section 8.1 [Fortran-interface routines], page 87), which reversed the dimensions for you. See Section 7.2 [Reversing array dimensions], page 78. Using ordinary Fortran array declarations like this works, but may yield suboptimal performance because the data may not be not aligned to exploit SIMD instructions on modern proessors (see Section 3.1 [SIMD alignment and fftw malloc], page 15). Better performance will often be obtained by allocating with fftw_alloc. See Section 7.5 [Allocating aligned memory in Fortran], page 82. Similar to the legacy Fortran interface (see Section 8.3 [FFTW Execution in Fortran], page 88), we currently recommend not using fftw_execute but rather using the more specialized functions like fftw_execute_dft (see Section 4.6 [New-array Execute Functions], page 38). However, you should execute the plan on the same arrays as the ones for which you created the plan, unless you are especially careful. See Section 7.4 [Plan execution in Fortran], page 81. To prevent you from using fftw_execute by mistake, the fftw3.f03 file does not provide an fftw_execute interface declaration. Multiple planner flags are combined with ior (equivalent to | in C). e.g. FFTW_MEASURE | FFTW_DESTROY_INPUT becomes ior(FFTW_MEASURE, FFTW_DESTROY_ INPUT). (You can also use + as long as you dont try to include a given flag more than once.)

7.1.1 Extended and quadruple precision in Fortran

If FFTW is compiled in long double (extended) precision (see Chapter 10 [Installation and Customization], page 97), you may be able to call the resulting fftwl_ routines (see Section 4.1.2 [Precision], page 21) from Fortran if your compiler supports the C_LONG_ DOUBLE_COMPLEX type code. Because some Fortran compilers do not support C_LONG_DOUBLE_COMPLEX, the fftwl_ declarations are segregated into a separate interface file fftw3l.f03, which you should include in addition to fftw3.f03 (which declares precision-independent FFTW_ constants): use, intrinsic :: iso_c_binding include fftw3.f03 include fftw3l.f03 We also support using the nonstandard __float128 quadruple-precision type provided by recent versions of gcc on 32- and 64-bit x86 hardware (see Chapter 10 [Installation and Customization], page 97), using the corresponding real(16) and complex(16) types supported by gfortran. The quadruple-precision fftwq_ functions (see Section 4.1.2 [Precision], page 21) are declared in a fftw3q.f03 interface file, which should be included in addition to fftw3l.f03, as above. You should also link with -lfftw3q -lquadmath -lm as in C.

7.2 Reversing array dimensions

A minor annoyance in calling FFTW from Fortran is that FFTWs array dimensions are defined in the C convention (row-major order), while Fortrans array dimensions are the opposite convention (column-major order). See Section 3.2 [Multi-dimensional Array Format], page 15. This is just a bookkeeping difference, with no effect on performance. The only

Chapter 7: Calling FFTW from Modern Fortran

79

consequence of this is that, whenever you create an FFTW plan for a multi-dimensional transform, you must always reverse the ordering of the dimensions. For example, consider the three-dimensional (L M N ) arrays: complex(C_DOUBLE_COMPLEX), dimension(L,M,N) :: in, out To plan a DFT for these arrays using fftw_plan_dft_3d, you could do: plan = fftw_plan_dft_3d(N,M,L, in,out, FFTW_FORWARD,FFTW_ESTIMATE) That is, from FFTWs perspective this is a N M L array. No data transposition need occur, as this is only notation. Similarly, to use the more generic routine fftw_plan_dft with the same arrays, you could do: integer(C_INT), dimension(3) :: n = [N,M,L] plan = fftw_plan_dft_3d(3, n, in,out, FFTW_FORWARD,FFTW_ESTIMATE) Note, by the way, that this is different from the legacy Fortran interface (see Section 8.1 [Fortran-interface routines], page 87), which automatically reverses the order of the array dimension for you. Here, you are calling the C interface directly, so there is no translation layer. An important thing to keep in mind is the implication of this for multidimensional realto-complex transforms (see Section 2.4 [Multi-Dimensional DFTs of Real Data], page 7). In C, a multidimensional real-to-complex DFT chops the last dimension roughly in half (N M L real input goes to N M L/2 + 1 complex output). In Fortran, because the array dimension notation is reversed, the first dimension of the complex data is chopped roughly in half. For example consider the r2c transform of L M N real input in Fortran: type(C_PTR) :: plan real(C_DOUBLE), dimension(L,M,N) :: in complex(C_DOUBLE_COMPLEX), dimension(L/2+1,M,N) :: out plan = fftw_plan_dft_r2c_3d(N,M,L, in,out, FFTW_ESTIMATE) ... call fftw_execute_dft_r2c(plan, in, out) Alternatively, for an in-place r2c transform, as described in the C documentation we must pad the first dimension of the real input with an extra two entries (which are ignored by FFTW) so as to leave enough space for the complex output. The input is allocated as a 2[L/2 + 1] M N array, even though only L M N of it is actually used. In this example, we will allocate the array as a pointer type, using fftw_alloc to ensure aligned memory for maximum performance (see Section 7.5 [Allocating aligned memory in Fortran], page 82); this also makes it easy to reference the same memory as both a real array and a complex array. real(C_DOUBLE), pointer :: in(:,:,:) complex(C_DOUBLE_COMPLEX), pointer :: out(:,:,:) type(C_PTR) :: plan, data data = fftw_alloc_complex(int((L/2+1) * M * N, C_SIZE_T)) call c_f_pointer(data, in, [2*(L/2+1),M,N]) call c_f_pointer(data, out, [L/2+1,M,N])

80

FFTW 3.3.3

plan ... call ... call call

= fftw_plan_dft_r2c_3d(N,M,L, in,out, FFTW_ESTIMATE) fftw_execute_dft_r2c(plan, in, out) fftw_destroy_plan(plan) fftw_free(data)

7.3 FFTW Fortran type reference

The following are the most important type correspondences between the C interface and Fortran: Plans (fftw_plan and variants) are type(C_PTR) (i.e. an opaque pointer). The C floating-point types double, float, and long double correspond to real(C_DOUBLE), real(C_FLOAT), and real(C_LONG_DOUBLE), respectively. The C complex types fftw_complex, fftwf_complex, and fftwl_complex correspond in Fortran to complex(C_DOUBLE_COMPLEX), complex(C_FLOAT_COMPLEX), and complex(C_LONG_DOUBLE_COMPLEX), respectively. Just as in C (see Section 4.1.2 [Precision], page 21), the FFTW subroutines and types are prefixed with fftw_, fftwf_, and fftwl_ for the different precisions, and link to different libraries (-lfftw3, -lfftw3f, and -lfftw3l on Unix), but use the same include file fftw3.f03 and the same constants (all of which begin with FFTW_). The exception is long double precision, for which you should also include fftw3l.f03 (see Section 7.1.1 [Extended and quadruple precision in Fortran], page 78). The C integer types int and unsigned (used for planner flags) become integer(C_ INT). The C integer type ptrdiff_t (e.g. in the Section 4.5.6 [64-bit Guru Interface], page 37) becomes integer(C_INTPTR_T), and size_t (in fftw_malloc etc.) becomes integer(C_SIZE_T). The fftw_r2r_kind type (see Section 4.3.6 [Real-to-Real Transform Kinds], page 30) becomes integer(C_FFTW_R2R_KIND). The various constant values of the C enumerated type (FFTW_R2HC etc.) become simply integer constants of the same names in Fortran. Numeric array pointer arguments (e.g. double *) become dimension(*), intent(out) arrays of the same type, or dimension(*), intent(in) if they are pointers to constant data (e.g. const int *). There are a few exceptions where numeric pointers refer to scalar outputs (e.g. for fftw_flops), in which case they are intent(out) scalar arguments in Fortran too. For the new-array execute functions (see Section 4.6 [New-array Execute Functions], page 38), the input arrays are declared dimension(*), intent(inout), since they can be modified in the case of in-place or FFTW_DESTROY_INPUT transforms. Pointer return values (e.g double *) become type(C_PTR). (If they are pointers to arrays, as for fftw_alloc_real, you can convert them back to Fortran array pointers with the standard intrinsic function c_f_pointer.) The fftw_iodim type in the guru interface (see Section 4.5.2 [Guru vector and transform sizes], page 34) becomes type(fftw_iodim) in Fortran, a derived data type (the Fortran analogue of Cs struct) with three integer(C_INT) components: n, is, and

Chapter 7: Calling FFTW from Modern Fortran

81

os, with the same meanings as in C. The fftw_iodim64 type in the 64-bit guru interface (see Section 4.5.6 [64-bit Guru Interface], page 37) is the same, except that its components are of type integer(C_INTPTR_T). Using the wisdom import/export functions from Fortran is a bit tricky, and is discussed in Section 7.6 [Accessing the wisdom API from Fortran], page 83. In brief, the FILE * arguments map to type(C_PTR), const char * to character(C_CHAR), dimension(*), intent(in) (null-terminated!), and the generic read-char/write-char functions map to type(C_FUNPTR). You may be wondering if you need to search-and-replace real(kind(0.0d0)) (or whatever your favorite Fortran spelling of double precision is) with real(C_DOUBLE) everywhere in your program, and similarly for complex and integer types. The answer is no; you can still use your existing types. As long as these types match their C counterparts, things should work without a hitch. The worst that can happen, e.g. in the (unlikely) event of a system where real(kind(0.0d0)) is different from real(C_DOUBLE), is that the compiler will give you a type-mismatch error. That is, if you dont use the iso_c_binding kinds you need to accept at least the theoretical possibility of having to change your code in response to compiler errors on some future machine, but you dont need to worry about silently compiling incorrect code that yields runtime errors.

7.4 Plan execution in Fortran

In C, in order to use a plan, one normally calls fftw_execute, which executes the plan to perform the transform on the input/output arrays passed when the plan was created (see Section 4.2 [Using Plans], page 22). The corresponding subroutine call in modern Fortran is: call fftw_execute(plan) However, we have had reports that this causes problems with some recent optimizing Fortran compilers. The problem is, because the input/output arrays are not passed as explicit arguments to fftw_execute, the semantics of Fortran (unlike C) allow the compiler to assume that the input/output arrays are not changed by fftw_execute. As a consequence, certain compilers end up repositioning the call to fftw_execute, assuming incorrectly that it does nothing to the arrays. There are various workarounds to this, but the safest and simplest thing is to not use fftw_execute in Fortran. Instead, use the functions described in Section 4.6 [New-array Execute Functions], page 38, which take the input/output arrays as explicit arguments. For example, if the plan is for a complex-data DFT and was created for the arrays in and out, you would do: call fftw_execute_dft(plan, in, out) There are a few things to be careful of, however: You must use the correct type of execute function, matching the way the plan was created. Complex DFT plans should use fftw_execute_dft, Real-input (r2c) DFT plans should use use fftw_execute_dft_r2c, and real-output (c2r) DFT plans should use fftw_execute_dft_c2r. The various r2r plans should use fftw_execute_r2r.

82

FFTW 3.3.3

Fortunately, if you use the wrong one you will get a compile-time type-mismatch error (unlike legacy Fortran). You should normally pass the same input/output arrays that were used when creating the plan. This is always safe. If you pass different input/output arrays compared to those used when creating the plan, you must abide by all the restrictions of the new-array execute functions (see Section 4.6 [New-array Execute Functions], page 38). The most tricky of these is the requirement that the new arrays have the same alignment as the original arrays; the best (and possibly only) way to guarantee this is to use the fftw_alloc functions to allocate your arrays (see Section 7.5 [Allocating aligned memory in Fortran], page 82). Alternatively, you can use the FFTW_UNALIGNED flag when creating the plan, in which case the plan does not depend on the alignment, but this may sacrifice substantial performance on architectures (like x86) with SIMD instructions (see Section 3.1 [SIMD alignment and fftw malloc], page 15).

7.5 Allocating aligned memory in Fortran

In order to obtain maximum performance in FFTW, you should store your data in arrays that have been specially aligned in memory (see Section 3.1 [SIMD alignment and fftw malloc], page 15). Enforcing alignment also permits you to safely use the new-array execute functions (see Section 4.6 [New-array Execute Functions], page 38) to apply a given plan to more than one pair of in/out arrays. Unfortunately, standard Fortran arrays do not provide any alignment guarantees. The only way to allocate aligned memory in standard Fortran is to allocate it with an external C function, like the fftw_alloc_real and fftw_alloc_complex functions. Fortunately, Fortran 2003 provides a simple way to associate such allocated memory with a standard Fortran array pointer that you can then use normally. We therefore recommend allocating all your input/output arrays using the following technique: 1. Declare a pointer, arr, to your array of the desired type and dimensions. For example, real(C_DOUBLE), pointer :: a(:,:) for a 2d real array, or complex(C_DOUBLE_ COMPLEX), pointer :: a(:,:,:) for a 3d complex array. 2. The number of elements to allocate must be an integer(C_SIZE_T). You can either declare a variable of this type, e.g. integer(C_SIZE_T) :: sz, to store the number of elements to allocate, or you can use the int(..., C_SIZE_T) intrinsic function. e.g. set sz = L * M * N or use int(L * M * N, C_SIZE_T) for an L M N array. 3. Declare a type(C_PTR) :: p to hold the return value from FFTWs allocation routine. Set p = fftw_alloc_real(sz) for a real array, or p = fftw_alloc_complex(sz) for a complex array. 4. Associate your pointer arr with the allocated memory p using the standard c_f_pointer subroutine: call c_f_pointer(p, arr, [...dimensions...]), where [...dimensions...]) are an array of the dimensions of the array (in the usual Fortran order). e.g. call c_f_pointer(p, arr, [L,M,N]) for an L M N array. (Alternatively, you can omit the dimensions argument if you specified the shape explicitly when declaring arr.) You can now use arr as a usual multidimensional array.

Chapter 7: Calling FFTW from Modern Fortran

83

5. When you are done using the array, deallocate the memory by call fftw_free(p) on p. For example, here is how we would allocate an L M 2d real array: real(C_DOUBLE), pointer :: arr(:,:) type(C_PTR) :: p p = fftw_alloc_real(int(L * M, C_SIZE_T)) call c_f_pointer(p, arr, [L,M]) ...use arr and arr(i,j) as usual... call fftw_free(p) and here is an L M N 3d complex array: complex(C_DOUBLE_COMPLEX), pointer :: arr(:,:,:) type(C_PTR) :: p p = fftw_alloc_complex(int(L * M * N, C_SIZE_T)) call c_f_pointer(p, arr, [L,M,N]) ...use arr and arr(i,j,k) as usual... call fftw_free(p) See Section 7.2 [Reversing array dimensions], page 78 for an example allocating a single array and associating both real and complex array pointers with it, for in-place real-to-complex transforms.

7.6 Accessing the wisdom API from Fortran

As explained in Section 3.3 [Words of Wisdom-Saving Plans], page 18, FFTW provides a wisdom API for saving plans to disk so that they can be recreated quickly. The C API for exporting (see Section 4.7.1 [Wisdom Export], page 40) and importing (see Section 4.7.2 [Wisdom Import], page 40) wisdom is somewhat tricky to use from Fortran, however, because of differences in file I/O and string types between C and Fortran.

7.6.1 Wisdom File Export/Import from Fortran

The easiest way to export and import wisdom is to do so using fftw_export_wisdom_ to_filename and fftw_wisdom_from_filename. The only trick is that these require you to pass a C string, which is an array of type CHARACTER(C_CHAR) that is terminated by C_NULL_CHAR. You can call them like this: integer(C_INT) :: ret ret = fftw_export_wisdom_to_filename(C_CHAR_my_wisdom.dat // C_NULL_CHAR) if (ret .eq. 0) stop error exporting wisdom to file ret = fftw_import_wisdom_from_filename(C_CHAR_my_wisdom.dat // C_NULL_CHAR) if (ret .eq. 0) stop error importing wisdom from file Note that prepending C_CHAR_ is needed to specify that the literal string is of kind C_CHAR, and we null-terminate the string by appending // C_NULL_CHAR. These functions return an integer(C_INT) (ret) which is 0 if an error occurred during export/import and nonzero otherwise. It is also possible to use the lower-level routines fftw_export_wisdom_to_file and fftw_ import_wisdom_from_file, which accept parameters of the C type FILE*, expressed in

84

FFTW 3.3.3

Fortran as type(C_PTR). However, you are then responsible for creating the FILE* yourself. You can do this by using iso_c_binding to define Fortran intefaces for the C library functions fopen and fclose, which is a bit strange in Fortran but workable.

7.6.2 Wisdom String Export/Import from Fortran

Dealing with FFTWs C string export/import is a bit more painful. In particular, the fftw_ export_wisdom_to_string function requires you to deal with a dynamically allocated C string. To get its length, you must define an interface to the C strlen function, and to deallocate it you must define an interface to C free: use, intrinsic :: iso_c_binding interface integer(C_INT) function strlen(s) bind(C, name=strlen) import type(C_PTR), value :: s end function strlen subroutine free(p) bind(C, name=free) import type(C_PTR), value :: p end subroutine free end interface Given these definitions, you can then export wisdom to a Fortran character array: character(C_CHAR), pointer :: s(:) integer(C_SIZE_T) :: slen type(C_PTR) :: p p = fftw_export_wisdom_to_string() if (.not. c_associated(p)) stop error exporting wisdom slen = strlen(p) call c_f_pointer(p, s, [slen+1]) ... call free(p) Note that slen is the length of the C string, but the length of the array is slen+1 because it includes the terminating null character. (You can omit the +1 if you dont want Fortran to know about the null character.) The standard c_associated function checks whether p is a null pointer, which is returned by fftw_export_wisdom_to_string if there was an error. To import wisdom from a string, use fftw_import_wisdom_from_string as usual; note that the argument of this function must be a character(C_CHAR) that is terminated by the C_NULL_CHAR character, like the s array above.

7.6.3 Wisdom Generic Export/Import from Fortran

The most generic wisdom export/import functions allow you to provide an arbitrary callback function to read/write one character at a time in any way you want. However, your callback function must be written in a special way, using the bind(C) attribute to be passed to a C interface.

Chapter 7: Calling FFTW from Modern Fortran

85

In particular, to call the generic wisdom export function fftw_export_wisdom, you would write a callback subroutine of the form: subroutine my_write_char(c, p) bind(C) use, intrinsic :: iso_c_binding character(C_CHAR), value :: c type(C_PTR), value :: p ...write c... end subroutine my_write_char Given such a subroutine (along with the corresponding interface definition), you could then export wisdom using: call fftw_export_wisdom(c_funloc(my_write_char), p) The standard c_funloc intrinsic converts a Fortran bind(C) subroutine into a C function pointer. The parameter p is a type(C_PTR) to any arbitrary data that you want to pass to my_write_char (or C_NULL_PTR if none). (Note that you can get a C pointer to Fortran data using the intrinsic c_loc, and convert it back to a Fortran pointer in my_write_char using c_f_pointer.) Similarly, to use the generic fftw_import_wisdom, you would define a callback function of the form: integer(C_INT) function my_read_char(p) bind(C) use, intrinsic :: iso_c_binding type(C_PTR), value :: p character :: c ...read a character c... my_read_char = ichar(c, C_INT) end function my_read_char .... integer(C_INT) :: ret ret = fftw_import_wisdom(c_funloc(my_read_char), p) if (ret .eq. 0) stop error importing wisdom Your function can return -1 if the end of the input is reached. Again, p is an arbitrary type(C_PTR that is passed through to your function. fftw_import_wisdom returns 0 if an error occurred and nonzero otherwise.

7.7 Defining an FFTW module

Rather than using the include statement to include the fftw3.f03 interface file in any subroutine where you want to use FFTW, you might prefer to define an FFTW Fortran module. FFTW does not install itself as a module, primarily because fftw3.f03 can be shared between different Fortran compilers while modules (in general) cannot. However, it is trivial to define your own FFTW module if you want. Just create a file containing: module FFTW3 use, intrinsic :: iso_c_binding

86

FFTW 3.3.3

include fftw3.f03 end module Compile this file into a module as usual for your compiler (e.g. with gfortran -c you will get a file fftw3.mod). Now, instead of include fftw3.f03, whenever you want to use FFTW routines you can just do: use FFTW3 as usual for Fortran modules. (You still need to link to the FFTW library, of course.)

Chapter 8: Calling FFTW from Legacy Fortran

87

8 Calling FFTW from Legacy Fortran

This chapter describes the interface to FFTW callable by Fortran code in older compilers not supporting the Fortran 2003 C interoperability features (see Chapter 7 [Calling FFTW from Modern Fortran], page 77). This interface has the major disadvantage that it is not type-checked, so if you mistake the argument types or ordering then your program will not have any compiler errors, and will likely crash at runtime. So, greater care is needed. Also, technically interfacing older Fortran versions to C is nonstandard, but in practice we have found that the techniques used in this chapter have worked with all known Fortran compilers for many years. The legacy Fortran interface differs from the C interface only in the prefix (dfftw_ instead of fftw_ in double precision) and a few other minor details. This Fortran interface is included in the FFTW libraries by default, unless a Fortran compiler isnt found on your system or --disable-fortran is included in the configure flags. We assume here that the reader is already familiar with the usage of FFTW in C, as described elsewhere in this manual. The MPI parallel interface to FFTW is not currently available to legacy Fortran.

8.1 Fortran-interface routines

Nearly all of the FFTW functions have Fortran-callable equivalents. The name of the legacy Fortran routine is the same as that of the corresponding C routine, but with the fftw_ prefix replaced by dfftw_.1 The single and long-double precision versions use sfftw_ and lfftw_, respectively, instead of fftwf_ and fftwl_; quadruple precision (real*16) is available on some systems as fftwq_ (see Section 4.1.2 [Precision], page 21). (Note that long double on x86 hardware is usually at most 80-bit extended precision, not quadruple precision.) For the most part, all of the arguments to the functions are the same, with the following exceptions: plan variables (what would be of type fftw_plan in C), must be declared as a type that is at least as big as a pointer (address) on your machine. We recommend using integer*8 everywhere, since this should always be big enough. Any function that returns a value (e.g. fftw_plan_dft) is converted into a subroutine. The return value is converted into an additional first parameter of this subroutine.2 The Fortran routines expect multi-dimensional arrays to be in column-major order, which is the ordinary format of Fortran arrays (see Section 3.2 [Multi-dimensional Array Format], page 15). They do this transparently and costlessly simply by reversing the order of the dimensions passed to FFTW, but this has one important consequence for multi-dimensional real-complex transforms, discussed below. Wisdom import and export is somewhat more tricky because one cannot easily pass files or strings between C and Fortran; see Section 8.5 [Wisdom of Fortran?], page 91.
1 2

Technically, Fortran 77 identifiers are not allowed to have more than 6 characters, nor may they contain underscores. Any compiler that enforces this limitation doesnt deserve to link to FFTW. The reason for this is that some Fortran implementations seem to have trouble with C function return values, and vice versa.

88

FFTW 3.3.3

Legacy Fortran cannot use the fftw_malloc dynamic-allocation routine. If you want to exploit the SIMD FFTW (see Section 3.1 [SIMD alignment and fftw malloc], page 15), youll need to figure out some other way to ensure that your arrays are at least 16-byte aligned. Since Fortran 77 does not have data structures, the fftw_iodim structure from the guru interface (see Section 4.5.2 [Guru vector and transform sizes], page 34) must be split into separate arguments. In particular, any fftw_iodim array arguments in the C guru interface become three integer array arguments (n, is, and os) in the Fortran guru interface, all of whose lengths should be equal to the corresponding rank argument. The guru planner interface in Fortran does not do any automatic translation between column-major and row-major; you are responsible for setting the strides etcetera to correspond to your Fortran arrays. However, as a slight bug that we are preserving for backwards compatibility, the plan_guru_r2r in Fortran does reverse the order of its kind array parameter, so the kind array of that routine should be in the reverse of the order of the iodim arrays (see above). In general, you should take care to use Fortran data types that correspond to (i.e. are the same size as) the C types used by FFTW. In practice, this correspondence is usually straightforward (i.e. integer corresponds to int, real corresponds to float, etcetera). The native Fortran double/single-precision complex type should be compatible with fftw_ complex/fftwf_complex. Such simple correspondences are assumed in the examples below.

8.2 FFTW Constants in Fortran

When creating plans in FFTW, a number of constants are used to specify options, such as FFTW_MEASURE or FFTW_ESTIMATE. The same constants must be used with the wrapper routines, but of course the C header files where the constants are defined cant be incorporated directly into Fortran code. Instead, we have placed Fortran equivalents of the FFTW constant definitions in the file fftw3.f, which can be found in the same directory as fftw3.h. If your Fortran compiler supports a preprocessor of some sort, you should be able to include or #include this file; otherwise, you can paste it directly into your code. In C, you combine different flags (like FFTW_PRESERVE_INPUT and FFTW_MEASURE) using the | operator; in Fortran you should just use +. (Take care not to add in the same flag more than once, though. Alternatively, you can use the ior intrinsic function standardized in Fortran 95.)

8.3 FFTW Execution in Fortran

In C, in order to use a plan, one normally calls fftw_execute, which executes the plan to perform the transform on the input/output arrays passed when the plan was created (see Section 4.2 [Using Plans], page 22). The corresponding subroutine call in legacy Fortran is: call dfftw_execute(plan) However, we have had reports that this causes problems with some recent optimizing Fortran compilers. The problem is, because the input/output arrays are not passed as explicit

Chapter 8: Calling FFTW from Legacy Fortran

89

arguments to dfftw_execute, the semantics of Fortran (unlike C) allow the compiler to assume that the input/output arrays are not changed by dfftw_execute. As a consequence, certain compilers end up optimizing out or repositioning the call to dfftw_execute, assuming incorrectly that it does nothing. There are various workarounds to this, but the safest and simplest thing is to not use dfftw_execute in Fortran. Instead, use the functions described in Section 4.6 [New-array Execute Functions], page 38, which take the input/output arrays as explicit arguments. For example, if the plan is for a complex-data DFT and was created for the arrays in and out, you would do: call dfftw_execute_dft(plan, in, out) There are a few things to be careful of, however: You must use the correct type of execute function, matching the way the plan was created. Complex DFT plans should use dfftw_execute_dft, Real-input (r2c) DFT plans should use use dfftw_execute_dft_r2c, and real-output (c2r) DFT plans should use dfftw_execute_dft_c2r. The various r2r plans should use dfftw_execute_r2r. You should normally pass the same input/output arrays that were used when creating the plan. This is always safe. If you pass different input/output arrays compared to those used when creating the plan, you must abide by all the restrictions of the new-array execute functions (see Section 4.6 [New-array Execute Functions], page 38). The most difficult of these, in Fortran, is the requirement that the new arrays have the same alignment as the original arrays, because there seems to be no way in legacy Fortran to obtain guaranteed-aligned arrays (analogous to fftw_malloc in C). You can, of course, use the FFTW_UNALIGNED flag when creating the plan, in which case the plan does not depend on the alignment, but this may sacrifice substantial performance on architectures (like x86) with SIMD instructions (see Section 3.1 [SIMD alignment and fftw malloc], page 15).

8.4 Fortran Examples

In C, you might have something like the following to transform a one-dimensional complex array: fftw_complex in[N], out[N]; fftw_plan plan; plan = fftw_plan_dft_1d(N,in,out,FFTW_FORWARD,FFTW_ESTIMATE); fftw_execute(plan); fftw_destroy_plan(plan); In Fortran, you would use the following to accomplish the same thing: double complex in, out dimension in(N), out(N) integer*8 plan call dfftw_plan_dft_1d(plan,N,in,out,FFTW_FORWARD,FFTW_ESTIMATE) call dfftw_execute_dft(plan, in, out)

90

FFTW 3.3.3

call dfftw_destroy_plan(plan) Notice how all routines are called as Fortran subroutines, and the plan is returned via the first argument to dfftw_plan_dft_1d. Notice also that we changed fftw_execute to dfftw_execute_dft (see Section 8.3 [FFTW Execution in Fortran], page 88). To do the same thing, but using 8 threads in parallel (see Chapter 5 [Multi-threaded FFTW], page 49), you would simply prefix these calls with: integer iret call dfftw_init_threads(iret) call dfftw_plan_with_nthreads(8) (You might want to check the value of iret: if it is zero, it indicates an unlikely error during thread initialization.) To transform a three-dimensional array in-place with C, you might do: fftw_complex arr[L][M][N]; fftw_plan plan; plan = fftw_plan_dft_3d(L,M,N, arr,arr, FFTW_FORWARD, FFTW_ESTIMATE); fftw_execute(plan); fftw_destroy_plan(plan); In Fortran, you would use this instead: double complex arr dimension arr(L,M,N) integer*8 plan call dfftw_plan_dft_3d(plan, L,M,N, arr,arr, & FFTW_FORWARD, FFTW_ESTIMATE) call dfftw_execute_dft(plan, arr, arr) call dfftw_destroy_plan(plan) Note that we pass the array dimensions in the natural order in both C and Fortran. To transform a one-dimensional real array in Fortran, you might do: double precision in dimension in(N) double complex out dimension out(N/2 + 1) integer*8 plan call dfftw_plan_dft_r2c_1d(plan,N,in,out,FFTW_ESTIMATE) call dfftw_execute_dft_r2c(plan, in, out) call dfftw_destroy_plan(plan) To transform a two-dimensional real array, out of place, you might use the following: double precision in dimension in(M,N)

Chapter 8: Calling FFTW from Legacy Fortran

91

double complex out dimension out(M/2 + 1, N) integer*8 plan call dfftw_plan_dft_r2c_2d(plan,M,N,in,out,FFTW_ESTIMATE) call dfftw_execute_dft_r2c(plan, in, out) call dfftw_destroy_plan(plan) Important: Notice that it is the first dimension of the complex output array that is cut in half in Fortran, rather than the last dimension as in C. This is a consequence of the interface routines reversing the order of the array dimensions passed to FFTW so that the Fortran program can use its ordinary column-major order.

8.5 Wisdom of Fortran?

In this section, we discuss how one can import/export FFTW wisdom (saved plans) to/from a Fortran program; we assume that the reader is already familiar with wisdom, as described in Section 3.3 [Words of Wisdom-Saving Plans], page 18. The basic problem is that is difficult to (portably) pass files and strings between Fortran and C, so we cannot provide a direct Fortran equivalent to the fftw_export_wisdom_ to_file, etcetera, functions. Fortran interfaces are provided for the functions that do not take file/string arguments, however: dfftw_import_system_wisdom, dfftw_import_ wisdom, dfftw_export_wisdom, and dfftw_forget_wisdom. So, for example, to import the system-wide wisdom, you would do: integer isuccess call dfftw_import_system_wisdom(isuccess) As usual, the C return value is turned into a first parameter; isuccess is non-zero on success and zero on failure (e.g. if there is no system wisdom installed). If you want to import/export wisdom from/to an arbitrary file or elsewhere, you can employ the generic dfftw_import_wisdom and dfftw_export_wisdom functions, for which you must supply a subroutine to read/write one character at a time. The FFTW package contains an example file doc/f77_wisdom.f demonstrating how to implement import_wisdom_ from_file and export_wisdom_to_file subroutines in this way. (These routines cannot be compiled into the FFTW library itself, lest all FFTW-using programs be required to link with the Fortran I/O library.)

Chapter 9: Upgrading from FFTW version 2

93

9 Upgrading from FFTW version 2

In this chapter, we outline the process for updating codes designed for the older FFTW 2 interface to work with FFTW 3. The interface for FFTW 3 is not backwards-compatible with the interface for FFTW 2 and earlier versions; codes written to use those versions will fail to link with FFTW 3. Nor is it possible to write compatibility wrappers to bridge the gap (at least not efficiently), because FFTW 3 has different semantics from previous versions. However, upgrading should be a straightforward process because the data formats are identical and the overall style of planning/execution is essentially the same. Unlike FFTW 2, there are no separate header files for real and complex transforms (or even for different precisions) in FFTW 3; all interfaces are defined in the <fftw3.h> header file.

Numeric Types
The main difference in data types is that fftw_complex in FFTW 2 was defined as a struct with macros c_re and c_im for accessing the real/imaginary parts. (This is binarycompatible with FFTW 3 on any machine except perhaps for some older Crays in single precision.) The equivalent macros for FFTW 3 are: #define c_re(c) ((c)[0]) #define c_im(c) ((c)[1]) This does not work if you are using the C99 complex type, however, unless you insert a double* typecast into the above macros (see Section 4.1.1 [Complex numbers], page 21). Also, FFTW 2 had an fftw_real typedef that was an alias for double (in double precision). In FFTW 3 you should just use double (or whatever precision you are employing).

Plans
The major difference between FFTW 2 and FFTW 3 is in the planning/execution division of labor. In FFTW 2, plans were found for a given transform size and type, and then could be applied to any arrays and for any multiplicity/stride parameters. In FFTW 3, you specify the particular arrays, stride parameters, etcetera when creating the plan, and the plan is then executed for those arrays (unless the guru interface is used) and those parameters only. (FFTW 2 had specific planner routines that planned for a particular array and stride, but the plan could still be used for other arrays and strides.) That is, much of the information that was formerly specified at execution time is now specified at planning time. Like FFTW 2s specific planner routines, the FFTW 3 planner overwrites the input/output arrays unless you use FFTW_ESTIMATE. FFTW 2 had separate data types fftw_plan, fftwnd_plan, rfftw_plan, and rfftwnd_ plan for complex and real one- and multi-dimensional transforms, and each type had its own destroy function. In FFTW 3, all plans are of type fftw_plan and all are destroyed by fftw_destroy_plan(plan). Where you formerly used fftw_create_plan and fftw_one to plan and compute a single 1d transform, you would now use fftw_plan_dft_1d to plan the transform. If you used

94

FFTW 3.3.3

the generic fftw function to execute the transform with multiplicity (howmany) and stride parameters, you would now use the advanced interface fftw_plan_many_dft to specify those parameters. The plans are now executed with fftw_execute(plan), which takes all of its parameters (including the input/output arrays) from the plan. In-place transforms no longer interpret their output argument as scratch space, nor is there an FFTW_IN_PLACE flag. You simply pass the same pointer for both the input and output arguments. (Previously, the output ostride and odist parameters were ignored for inplace transforms; now, if they are specified via the advanced interface, they are significant even in the in-place case, although they should normally equal the corresponding input parameters.) The FFTW_ESTIMATE and FFTW_MEASURE flags have the same meaning as before, although the planning time will differ. You may also consider using FFTW_PATIENT, which is like FFTW_ MEASURE except that it takes more time in order to consider a wider variety of algorithms. For multi-dimensional complex DFTs, instead of fftwnd_create_plan (or fftw2d_create_ plan or fftw3d_create_plan), followed by fftwnd_one, you would use fftw_plan_dft (or fftw_plan_dft_2d or fftw_plan_dft_3d). followed by fftw_execute. If you used fftwnd to to specify strides etcetera, you would instead specify these via fftw_plan_many_dft. The analogues to rfftw_create_plan and rfftw_one with FFTW_REAL_TO_COMPLEX or FFTW_COMPLEX_TO_REAL directions are fftw_plan_r2r_1d with kind FFTW_R2HC or FFTW_ HC2R, followed by fftw_execute. The stride etcetera arguments of rfftw are now in fftw_ plan_many_r2r. Instead of rfftwnd_create_plan (or rfftw2d_create_plan or rfftw3d_create_plan) followed by rfftwnd_one_real_to_complex or rfftwnd_one_complex_to_real, you now use fftw_plan_dft_r2c (or fftw_plan_dft_r2c_2d or fftw_plan_dft_r2c_3d) or fftw_ plan_dft_c2r (or fftw_plan_dft_c2r_2d or fftw_plan_dft_c2r_3d), respectively, followed by fftw_execute. As usual, the strides etcetera of rfftwnd_real_to_complex or rfftwnd_complex_to_real are no specified in the advanced planner routines, fftw_plan_ many_dft_r2c or fftw_plan_many_dft_c2r.

Wisdom
In FFTW 2, you had to supply the FFTW_USE_WISDOM flag in order to use wisdom; in FFTW 3, wisdom is always used. (You could simulate the FFTW 2 wisdom-less behavior by calling fftw_forget_wisdom after every planner call.) The FFTW 3 wisdom import/export routines are almost the same as before (although the storage format is entirely different). There is one significant difference, however. In FFTW 2, the import routines would never read past the end of the wisdom, so you could store extra data beyond the wisdom in the same file, for example. In FFTW 3, the file-import routine may read up to a few hundred bytes past the end of the wisdom, so you cannot store other data just beyond it.1 Wisdom has been enhanced by additional humility in FFTW 3: whereas FFTW 2 would re-use wisdom for a given transform size regardless of the stride etc., in FFTW 3 wisdom is
1

We do our own buffering because GNU libc I/O routines are horribly slow for single-character I/O, apparently for thread-safety reasons (whether you are using threads or not).

Chapter 9: Upgrading from FFTW version 2

95

only used with the strides etc. for which it was created. Unfortunately, this means FFTW 3 has to create new plans from scratch more often than FFTW 2 (in FFTW 2, planning e.g. one transform of size 1024 also created wisdom for all smaller powers of 2, but this no longer occurs). FFTW 3 also has the new routine fftw_import_system_wisdom to import wisdom from a standard system-wide location.

Memory allocation
In FFTW 3, we recommend allocating your arrays with fftw_malloc and deallocating them with fftw_free; this is not required, but allows optimal performance when SIMD acceleration is used. (Those two functions actually existed in FFTW 2, and worked the same way, but were not documented.) In FFTW 2, there were fftw_malloc_hook and fftw_free_hook functions that allowed the user to replace FFTWs memory-allocation routines (e.g. to implement different errorhandling, since by default FFTW prints an error message and calls exit to abort the program if malloc returns NULL). These hooks are not supported in FFTW 3; those few users who require this functionality can just directly modify the memory-allocation routines in FFTW (they are defined in kernel/alloc.c).

Fortran interface
In FFTW 2, the subroutine names were obtained by replacing fftw_ with fftw_f77; in FFTW 3, you replace fftw_ with dfftw_ (or sfftw_ or lfftw_, depending upon the precision). In FFTW 3, we have begun recommending that you always declare the type used to store plans as integer*8. (Too many people didnt notice our instruction to switch from integer to integer*8 for 64-bit machines.) In FFTW 3, we provide a fftw3.f header file to include in your code (and which is officially installed on Unix systems). (In FFTW 2, we supplied a fftw_f77.i file, but it was not installed.) Otherwise, the C-Fortran interface relationship is much the same as it was before (e.g. return values become initial parameters, and multi-dimensional arrays are in column-major order). Unlike FFTW 2, we do provide some support for wisdom import/export in Fortran (see Section 8.5 [Wisdom of Fortran?], page 91).

Threads
Like FFTW 2, only the execution routines are thread-safe. All planner routines, etcetera, should be called by only a single thread at a time (see Section 5.4 [Thread safety], page 51). Unlike FFTW 2, there is no special FFTW_THREADSAFE flag for the planner to allow a given plan to be usable by multiple threads in parallel; this is now the case by default. The multi-threaded version of FFTW 2 required you to pass the number of threads each time you execute the transform. The number of threads is now stored in the plan, and is specified before the planner is called by fftw_plan_with_nthreads. The threads initialization routine used to be called fftw_threads_init and would return zero on success;

96

FFTW 3.3.3

the new routine is called fftw_init_threads and returns zero on failure. See Chapter 5 [Multi-threaded FFTW], page 49. There is no separate threads header file in FFTW 3; all the function prototypes are in <fftw3.h>. However, you still have to link to a separate library (-lfftw3_threads lfftw3 -lm on Unix), as well as to the threading library (e.g. POSIX threads on Unix).

Chapter 10: Installation and Customization

97

10 Installation and Customization

This chapter describes the installation and customization of FFTW, the latest version of which may be downloaded from the FFTW home page. In principle, FFTW should work on any system with an ANSI C compiler (gcc is fine). However, planner time is drastically reduced if FFTW can exploit a hardware cycle counter; FFTW comes with cycle-counter support for all modern general-purpose CPUs, but you may need to add a couple of lines of code if your compiler is not yet supported (see Section 10.3 [Cycle Counters], page 100). (On Unix, there will be a warning at the end of the configure output if no cycle counter is found.) Installation of FFTW is simplest if you have a Unix or a GNU system, such as GNU/Linux, and we describe this case in the first section below, including the use of special configuration options to e.g. install different precisions or exploit optimizations for particular architectures (e.g. SIMD). Compilation on non-Unix systems is a more manual process, but we outline the procedure in the second section. It is also likely that pre-compiled binaries will be available for popular systems. Finally, we describe how you can customize FFTW for particular needs by generating codelets for fast transforms of sizes not supported efficiently by the standard FFTW distribution.

10.1 Installation on Unix

FFTW comes with a configure program in the GNU style. Installation can be as simple as: ./configure make make install This will build the uniprocessor complex and real transform libraries along with the test programs. (We recommend that you use GNU make if it is available; on some systems it is called gmake.) The make install command installs the fftw and rfftw libraries in standard places, and typically requires root privileges (unless you specify a different install directory with the --prefix flag to configure). You can also type make check to put the FFTW test programs through their paces. If you have problems during configuration or compilation, you may want to run make distclean before trying again; this ensures that you dont have any stale files left over from previous compilation attempts. The configure script chooses the gcc compiler by default, if it is available; you can select some other compiler with: ./configure CC="<the name of your C compiler>" The configure script knows good CFLAGS (C compiler flags) for a few systems. If your system is not known, the configure script will print out a warning. In this case, you should re-configure FFTW with the command ./configure CFLAGS="<write your CFLAGS here>"

98

FFTW 3.3.3

and then compile as usual. If you do find an optimal set of CFLAGS for your system, please let us know what they are (along with the output of config.guess) so that we can include them in future releases. configure supports all the standard flags defined by the GNU Coding Standards; see the INSTALL file in FFTW or the GNU web page. Note especially --help to list all flags and --enable-shared to create shared, rather than static, libraries. configure also accepts a few FFTW-specific flags, particularly: --enable-float: Produces a single-precision version of FFTW (float) instead of the default double-precision (double). See Section 4.1.2 [Precision], page 21. --enable-long-double: Produces a long-double precision version of FFTW (long double) instead of the default double-precision (double). The configure script will halt with an error message if long double is the same size as double on your machine/compiler. See Section 4.1.2 [Precision], page 21. --enable-quad-precision: Produces a quadruple-precision version of FFTW using the nonstandard __float128 type provided by gcc 4.6 or later on x86, x86-64, and Itanium architectures, instead of the default double-precision (double). The configure script will halt with an error message if the compiler is not gcc version 4.6 or later or if gccs libquadmath library is not installed. See Section 4.1.2 [Precision], page 21. --enable-threads: Enables compilation and installation of the FFTW threads library (see Chapter 5 [Multi-threaded FFTW], page 49), which provides a simple interface to parallel transforms for SMP systems. By default, the threads routines are not compiled. --enable-openmp: Like --enable-threads, but using OpenMP compiler directives in order to induce parallelism rather than spawning its own threads directly, and installing an fftw3_omp library rather than an fftw3_threads library (see Chapter 5 [Multi-threaded FFTW], page 49). You can use both --enable-openmp and --enablethreads since they compile/install libraries with different names. By default, the OpenMP routines are not compiled. --with-combined-threads: By default, if --enable-threads is used, the threads support is compiled into a separate library that must be linked in addition to the main FFTW library. This is so that users of the serial library do not need to link the system threads libraries. If --with-combined-threads is specified, however, then no separate threads library is created, and threads are included in the main FFTW library. This is mainly useful under Windows, where no system threads library is required and inter-library dependencies are problematic. --enable-mpi: Enables compilation and installation of the FFTW MPI library (see Chapter 6 [Distributed-memory FFTW with MPI], page 53), which provides parallel transforms for distributed-memory systems with MPI. (By default, the MPI routines are not compiled.) See Section 6.1 [FFTW MPI Installation], page 53. --disable-fortran: Disables inclusion of legacy-Fortran wrapper routines (see Chapter 8 [Calling FFTW from Legacy Fortran], page 87) in the standard FFTW libraries. These wrapper routines increase the library size by only a negligible amount, so they are included by default as long as the configure script finds a Fortran compiler on your system. (To specify a particular Fortran compiler foo, pass F77=foo to configure.)

Chapter 10: Installation and Customization

99

--with-g77-wrappers: By default, when Fortran wrappers are included, the wrappers employ the linking conventions of the Fortran compiler detected by the configure script. If this compiler is GNU g77, however, then two versions of the wrappers are included: one with g77s idiosyncratic convention of appending two underscores to identifiers, and one with the more common convention of appending only a single underscore. This way, the same FFTW library will work with both g77 and other Fortran compilers, such as GNU gfortran. However, the converse is not true: if you configure with a different compiler, then the g77-compatible wrappers are not included. By specifying --with-g77-wrappers, the g77-compatible wrappers are included in addition to wrappers for whatever Fortran compiler configure finds. --with-slow-timer: Disables the use of hardware cycle counters, and falls back on gettimeofday or clock. This greatly worsens performance, and should generally not be used (unless you dont have a cycle counter but still really want an optimized plan regardless of the time). See Section 10.3 [Cycle Counters], page 100. --enable-sse, --enable-sse2, --enable-avx, --enable-altivec, --enable-neon: Enable the compilation of SIMD code for SSE (Pentium III+), SSE2 (Pentium IV+), AVX (Sandy Bridge, Interlagos), AltiVec (PowerPC G4+), NEON (some ARM processors). SSE, AltiVec, and NEON only work with --enable-float (above). SSE2 works in both single and double precision (and is simply SSE in single precision). The resulting code will still work on earlier CPUs lacking the SIMD extensions (SIMD is automatically disabled, although the FFTW library is still larger). These options require a compiler supporting SIMD extensions, and compiler support is always a bit flaky: see the FFTW FAQ for a list of compiler versions that have problems compiling FFTW. With AltiVec and gcc, you may have to use the -mabi=altivec option when compiling any code that links to FFTW, in order to properly align the stack; otherwise, FFTW could crash when it tries to use an AltiVec feature. (This is not necessary on MacOS X.) With SSE/SSE2 and gcc, you should use a version of gcc that properly aligns the stack when compiling any code that links to FFTW. By default, gcc 2.95 and later versions align the stack as needed, but you should not compile FFTW with the -Os option or the -mpreferred-stack-boundary option with an argument less than 4. Because of the large variety of ARM processors and ABIs, FFTW does not attempt to guess the correct gcc flags for generating NEON code. In general, you will have to provide them on the command line. This command line is known to have worked at least once: ./configure --with-slow-timer --host=arm-linux-gnueabi \ --enable-single --enable-neon \ "CC=arm-linux-gnueabi-gcc -march=armv7-a -mfloat-abi=softfp" To force configure to use a particular C compiler foo (instead of the default, usually gcc), pass CC=foo to the configure script; you may also need to set the flags via the variable CFLAGS as described above.

100

FFTW 3.3.3

10.2 Installation on non-Unix systems

It should be relatively straightforward to compile FFTW even on non-Unix systems lacking the niceties of a configure script. Basically, you need to edit the config.h header (copy it from config.h.in) to #define the various options and compiler characteristics, and then compile all the .c files in the relevant directories. The config.h header contains about 100 options to set, each one initially an #undef, each documented with a comment, and most of them fairly obvious. For most of the options, you should simply #define them to 1 if they are applicable, although a few options require a particular value (e.g. SIZEOF_LONG_LONG should be defined to the size of the long long type, in bytes, or zero if it is not supported). We will likely post some sample config.h files for various operating systems and compilers for you to use (at least as a starting point). Please let us know if you have to hand-create a configuration file (and/or a pre-compiled binary) that you want to share. To create the FFTW library, you will then need to compile all of the .c files in the kernel, dft, dft/scalar, dft/scalar/codelets, rdft, rdft/scalar, rdft/scalar/r2cf, rdft/scalar/r2cb, rdft/scalar/r2r, reodft, and api directories. If you are compiling with SIMD support (e.g. you defined HAVE_SSE2 in config.h), then you also need to compile the .c files in the simd-support, {dft,rdft}/simd, {dft,rdft}/simd/* directories. Once these files are all compiled, link them into a library, or a shared library, or directly into your program. To compile the FFTW test program, additionally compile the code in the libbench2/ directory, and link it into a library. Then compile the code in the tests/ directory and link it to the libbench2 and FFTW libraries. To compile the fftw-wisdom (command-line) tool (see Section 4.7.4 [Wisdom Utilities], page 41), compile tools/fftw-wisdom.c and link it to the libbench2 and FFTW libraries

10.3 Cycle Counters

FFTWs planner actually executes and times different possible FFT algorithms in order to pick the fastest plan for a given n. In order to do this in as short a time as possible, however, the timer must have a very high resolution, and to accomplish this we employ the hardware cycle counters that are available on most CPUs. Currently, FFTW supports the cycle counters on x86, PowerPC/POWER, Alpha, UltraSPARC (SPARC v9), IA64, PA-RISC, and MIPS processors. Access to the cycle counters, unfortunately, is a compiler and/or operating-system dependent task, often requiring inline assembly language, and it may be that your compiler is not supported. If you are not supported, FFTW will by default fall back on its estimator (effectively using FFTW_ESTIMATE for all plans). You can add support by editing the file kernel/cycle.h; normally, this will involve adapting one of the examples already present in order to use the inline-assembler syntax for your C compiler, and will only require a couple of lines of code. Anyone adding support for a new system to cycle.h is encouraged to email us at fftw@fftw.org. If a cycle counter is not available on your system (e.g. some embedded processor), and you dont want to use estimated plans, as a last resort you can use the --with-slow-timer

Chapter 10: Installation and Customization

101

option to configure (on Unix) or #define WITH_SLOW_TIMER in config.h (elsewhere). This will use the much lower-resolution gettimeofday function, or even clock if the former is unavailable, and planning will be extremely slow.

10.4 Generating your own code

The directory genfft contains the programs that were used to generate FFTWs codelets, which are hard-coded transforms of small sizes. We do not expect casual users to employ the generator, which is a rather sophisticated program that generates directed acyclic graphs of FFT algorithms and performs algebraic simplifications on them. It was written in Objective Caml, a dialect of ML, which is available at http://caml.inria.fr/ocaml/index.en. html. If you have Objective Caml installed (along with recent versions of GNU autoconf, automake, and libtool), then you can change the set of codelets that are generated or play with the generation options. The set of generated codelets is specified by the {dft,rdft}/{codelets,simd}/*/Makefile.am files. For example, you can add efficient REDFT codelets of small sizes by modifying rdft/codelets/r2r/Makefile.am. After you modify any Makefile.am files, you can type sh bootstrap.sh in the top-level directory followed by make to re-generate the files. We do not provide more details about the code-generation process, since we do not expect that most users will need to generate their own code. However, feel free to contact us at fftw@fftw.org if you are interested in the subject. You might find it interesting to learn Caml and/or some modern programming techniques that we used in the generator (including monadic programming), especially if you heard the rumor that Java and object-oriented programming are the latest advancement in the field. The internal operation of the codelet generator is described in the paper, A Fast Fourier Transform Compiler, by M. Frigo, which is available from the FFTW home page and also appeared in the Proceedings of the 1999 ACM SIGPLAN Conference on Programming Language Design and Implementation (PLDI).

Chapter 11: Acknowledgments

103

11 Acknowledgments
Matteo Frigo was supported in part by the Special Research Program SFB F011 AURORA of the Austrian Science Fund FWF and by MIT Lincoln Laboratory. For previous versions of FFTW, he was supported in part by the Defense Advanced Research Projects Agency (DARPA), under Grants N00014-94-1-0985 and F30602-97-1-0270, and by a Digital Equipment Corporation Fellowship. Steven G. Johnson was supported in part by a Dept. of Defense NDSEG Fellowship, an MIT Karl Taylor Compton Fellowship, and by the Materials Research Science and Engineering Center program of the National Science Foundation under award DMR-9400334. Code for the Cell Broadband Engine was graciously donated to the FFTW project by the IBM Austin Research Lab and included in fftw-3.2. (This code was removed in fftw-3.3.) Code for the MIPS paired-single SIMD support was graciously donated to the FFTW project by CodeSourcery, Inc. We are grateful to Sun Microsystems Inc. for its donation of a cluster of 9 8-processor Ultra HPC 5000 SMPs (24 Gflops peak). These machines served as the primary platform for the development of early versions of FFTW. We thank Intel Corporation for donating a four-processor Pentium Pro machine. We thank the GNU/Linux community for giving us a decent OS to run on that machine. We are thankful to the AMD corporation for donating an AMD Athlon XP 1700+ computer to the FFTW project. We thank the Compaq/HP testdrive program and VA Software Corporation (SourceForge.net) for providing remote access to machines that were used to test FFTW. The genfft suite of code generators was written using Objective Caml, a dialect of ML. Objective Caml is a small and elegant language developed by Xavier Leroy. The implementation is available from http://caml.inria.fr/. In previous releases of FFTW, genfft was written in Caml Light, by the same authors. An even earlier implementation of genfft was written in Scheme, but Caml is definitely better for this kind of application. FFTW uses many tools from the GNU project, including automake, texinfo, and libtool. Prof. Charles E. Leiserson of MIT provided continuous support and encouragement. This program would not exist without him. Charles also proposed the name codelets for the basic FFT blocks. Prof. John D. Joannopoulos of MIT demonstrated continuing tolerance of Stevens extracurricular computer-science activities, as well as remarkable creativity in working them into his grant proposals. Stevens physics degree would not exist without him. Franz Franchetti wrote SIMD extensions to FFTW 2, which eventually led to the SIMD support in FFTW 3. Stefan Kral wrote most of the K7 code generator distributed with FFTW 3.0.x and 3.1.x. Andrew Sterian contributed the Windows timing code in FFTW 2.

104

FFTW 3.3.3

Didier Miras reported a bug in the test procedure used in FFTW 1.2. We now use a completely different test algorithm by Funda Ergun that does not require a separate FFT program to compare against. Wolfgang Reimer contributed the Pentium cycle counter and a few fixes that help portability. Ming-Chang Liu uncovered a well-hidden bug in the complex transforms of FFTW 2.0 and supplied a patch to correct it. The FFTW FAQ was written in bfnn (Bizarre Format With No Name) and formatted using the tools developed by Ian Jackson for the Linux FAQ. We are especially thankful to all of our users for their continuing support, feedback, and interest during our development of FFTW.

Chapter 12: License and Copyright

105

12 License and Copyright

FFTW is Copyright c 2003, 2007-11 Matteo Frigo, Copyright c 2003, 2007-11 Massachusetts Institute of Technology. FFTW is free software; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation; either version 2 of the License, or (at your option) any later version. This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details. You should have received a copy of the GNU General Public License along with this program; if not, write to the Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA You can also find the GPL on the GNU web site. In addition, we kindly ask you to acknowledge FFTW and its authors in any program or publication in which you use FFTW. (You are not required to do so; it is up to your common sense to decide whether you want to comply with this request or not.) For general publications, we suggest referencing: Matteo Frigo and Steven G. Johnson, The design and implementation of FFTW3, Proc. IEEE 93 (2), 216231 (2005). Non-free versions of FFTW are available under terms different from those of the General Public License. (e.g. they do not require you to accompany any object code using FFTW with the corresponding source code.) For these alternative terms you must purchase a license from MITs Technology Licensing Office. Users interested in such a license should contact us (fftw@fftw.org) for more information.

Chapter 13: Concept Index

107

13 Concept Index
6
64-bit architecture . . . . . . . . . . . . . . . . . . . . . . 37, 55, 80

F
FFTW . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 fftw-wisdom utility . . . . . . . . . . . . . . . . . . . . . . . . . 19, 41 fftw-wisdom-to-conf utility . . . . . . . . . . . . . . . . . . . . . 41 flags . . . . . . . . . . . . . . . . 4, 7, 25, 28, 30, 35, 37, 78, 88 Fortran interface . . . . . . . . . . . . . . . . . . . . 16, 74, 77, 87 Fortran-callable wrappers . . . . . . . . . . . . . . . . . . . . . . 98 frequency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4, 42

A
advanced interface . . . . . . . . . 1, 6, 16, 31, 57, 69, 71 algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 alignment . . . . . . . . . . . . . . . . . . . 22, 26, 39, 68, 78, 82 AltiVec . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 AVX . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

G
g77 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99 guru interface . . . . . . . . . . . . . . . . . . . . . 1, 6, 34, 80, 88

B
basic interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1, 3, 24 block distribution . . . . . . . . . . . . . . . . . . . . . . 56, 57, 66

H
halfcomplex format . . . . . . . . . . . . . . . . . . . . . . 7, 11, hc2r . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11, HDF5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Hermitian . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6, howmany loop . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . howmany parameter . . . . . . . . . . . . . . . . . . . . . . . . . . . 42 26 56 42 35 32

C
C multi-dimensional arrays . . . . . . . . . . . . . . . . . . . . 16 C++ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5, 15, 17, 21, 22 c2r . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7, 26, 28 C99 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17, 21, 22 Caml . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101, 103 code generator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2, 101 codelet . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2, 97, 101, 103 collective function . . . . . . . . . . . . . . . . . . 55, 65, 68, 70 column-major . . . . . . . . . . . . . . . . . . . . . . 16, 78, 87, 91 compiler . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2, 97, 99, 100 compiler flags . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97, 99 configuration routines . . . . . . . . . . . . . . . . . . . . . . . . . . 41 configure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49, 53, 97 cycle counter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97, 100

I
IDCT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12, 31, in-place . . . . . . . . . 4, 7, 25, 28, 29, 30, 37, 64, 79, installation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . interleaved format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iso c binding . . . . . . . . . . . . . . . . . . . . . . . . . . . 74, 77, 44 80 97 34 78

K
kind (r2r) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10, 30

D
data distribution . . . . . . . . . . . . 53, 55, 56, 60, 63, 69 DCT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11, 30, 43 deadlock . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65 Devil . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 DFT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1, 4, 42 DHT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13, 45 discrete cosine transform . . . . . . . . . . . . . . . 11, 30, 43 discrete Fourier transform . . . . . . . . . . . . . . . . . . . 1, 42 discrete Hartley transform . . . . . . . . . . . . . . 13, 30, 45 discrete sine transform . . . . . . . . . . . . . . . . . 11, 31, 44 dist . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32, 35 DST . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11, 31, 44

L
linking on Unix . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50, 54 LISP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103 load balancing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58, 66

M
MIPS PS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 monadic programming . . . . . . . . . . . . . . . . . . . . . . . . 101 MPI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53, 98 MPI communicator . . . . . . . . . . . . . . . . . 53, 68, 70, 74 MPI I/O . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56, 64 mpicc . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53, 54

E
Ecclesiastes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 execute . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1, 4, 38

N
new-array execution . . . . . . . . . . . . . . . . 38, 68, 73, 75

108

FFTW 3.3.3

normalization . . . 4, 9, 11, 12, 13, 25, 28, 30, 42, 43, 44, 45 number of threads . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51

REDFT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11, 43, 101 RODFT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11, 44 row-major . . . . . . . . . . . . . . . 15, 25, 30, 35, 57, 61, 78

O
OpenMP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49, 50, 51 out-of-place . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26, 29

S
saving plans to disk . . . . . . . . . . . . . . . . . 18, 40, 64, shared-memory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . SIMD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3, 15, split format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . SSE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . SSE2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . stride . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16, 32, 34, 83 49 78 34 15 15 71

P
padding . . . . . . . . . . . . . . . . . . . . . . . 6, 8, 28, 29, 60, 79 parallel transform . . . . . . . . . . . . . . . . . . . . . . . . . . 49, 53 partial order . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 plan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1, 4 planner . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 portability . . . . 15, 19, 21, 49, 77, 81, 87, 88, 91, 97 precision . . . . . . . . 5, 7, 15, 21, 22, 54, 67, 78, 80, 98

T
thread safety . . . . . . . . . . . . . . . . . . . . . . . . . . . 50, 51, 67 threads . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49, 51, 66, 98 transpose . . . . . . . . . . . . . . . . . . . . 58, 61, 62, 66, 67, 73

R
r2c . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7, 11, 27, 46, 71 r2c/c2r multi-dimensional array format . . 8, 28, 79, 91 r2hc . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 r2r . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10, 29, 42, 62, 73 rank . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 real-even DFT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11, 43 real-odd DFT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11, 44

V
vector . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

W
wisdom . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18, 40, 64, 83 wisdom, problems with . . . . . . . . . . . . . . . . . . . . . . . . 18 wisdom, system-wide . . . . . . . . . . . . . . . . . . . . . . . 19, 41

Chapter 14: Library Index

109

14 Library Index
C
c_associated . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . C_DOUBLE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77, C_DOUBLE_COMPLEX . . . . . . . . . . . . . . . . . . . . . . . . . 77, c_f_pointer . . . . . . . . . . . . . . . . . . . . 79, 80, 82, 84, C_FFTW_R2R_KIND . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . C_FLOAT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . C_FLOAT_COMPLEX . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . c_funloc . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . C_FUNPTR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . C_INT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77, C_INTPTR_T . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . c_loc . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . C_LONG_DOUBLE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . C_LONG_DOUBLE_COMPLEX . . . . . . . . . . . . . . . . . . . . . . . C_PTR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . C_SIZE_T . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84 80 80 85 80 80 80 85 81 80 80 85 80 80 77 80 fftw_execute_split_dft_c2r . . . . . . . . . . . . . . . . . fftw_execute_split_dft_r2c . . . . . . . . . . . . . . . . . FFTW_EXHAUSTIVE . . . . . . . . . . . . . . . . . . . . . . . . . . 18, fftw_export_wisdom . . . . . . . . . . . . . . . . . . . . . . . 40, fftw_export_wisdom_to_file . . . . . . . . . . . . . . . . . fftw_export_wisdom_to_filename . . . . . 18, 40, fftw_export_wisdom_to_string . . . . . . . . . . . 40, fftw_flops . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23, 66, fftw_forget_wisdom . . . . . . . . . . . . . . . . . . . . . . . 18, FFTW_FORWARD . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4, 7, fftw_fprint_plan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . fftw_free . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4, 15, FFTW_HC2R . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11, fftw_import wisdom_from_filename . . . . . . . . . . . fftw_import_system_wisdom . . . . . . . . . . . . . . . 19, fftw_import_wisdom . . . . . . . . . . . . . . . . . . . . . . . 40, fftw_import_wisdom_from_file . . . . . . . . . . . . . . . fftw_import_wisdom_from_filename . . . . . . . 18, fftw_import_wisdom_from_string . . . . . . . . . 40, fftw_init_threads . . . . . . . . . . . . . . . . . 50, 54, 67, fftw_iodim . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34, 80, fftw_iodim64 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38, fftw_malloc . . . . . . . . . . . . . . 3, 15, 17, 22, 57, 68, FFTW_MEASURE . . . . . . . . . . . . . . . . . . . . . . . . 4, 18, 26, fftw_mpi_broadcast_wisdom . . . . . . . . . . . . . . . 65, fftw_mpi_cleanup . . . . . . . . . . . . . . . . . . . . . . . . . 54, FFTW_MPI_DEFAULT_BLOCK . . . . . . . . . . . . . . . 57, 63, fftw_mpi_execute_dft . . . . . . . . . . . . . . . . . . . . 68, fftw_mpi_execute_dft_c2r . . . . . . . . . . . . . . . . . . . fftw_mpi_execute_dft_r2c . . . . . . . . . . . . . . . . . . . fftw_mpi_execute_r2r . . . . . . . . . . . . . . . . . . . . 68, fftw_mpi_gather_wisdom . . . . . . . . . . . . . . . . . . 65, fftw_mpi_init . . . . . . . . . . . . . . . . . . 54, 55, 65, 67, fftw_mpi_local_size . . . . . . . . . . . . . . . . . . . . . . . . . fftw_mpi_local_size_1d . . . . . . . . . . . . . . . . . . 59, fftw_mpi_local_size_2d . . . . . . . . . . . . . . . 55, 57, fftw_mpi_local_size_2d_transposed . . . . . 63, fftw_mpi_local_size_3d . . . . . . . . . . . . . . . . . . . . . . fftw_mpi_local_size_3d_transposed . . . . . 59, fftw_mpi_local_size_many . . . . . . . . . . . . . . . . 57, fftw_mpi_local_size_many_1d . . . . . . . . . . . . . . . . fftw_mpi_local_size_many_transposed . . . 63, fftw_mpi_local_size_transposed . . . . . . . . . . . . . fftw_mpi_plan_dft . . . . . . . . . . . . . . . . . . . . . . . . . . . . fftw_mpi_plan_dft_1d . . . . . . . . . . . . . . . . . . . . . . . . fftw_mpi_plan_dft_2d . . . . . . . . . . . . . . . . . . . . 55, fftw_mpi_plan_dft_3d . . . . . . . . . . . . . . . . . . . . . . . . fftw_mpi_plan_dft_c2r . . . . . . . . . . . . . . . . . . . . . . . fftw_mpi_plan_dft_c2r_2d . . . . . . . . . . . . . . . . . . . fftw_mpi_plan_dft_c2r_3d . . . . . . . . . . . . . . . . . . . fftw_mpi_plan_dft_r2c . . . . . . . . . . . . . . . . . . . . . . . fftw_mpi_plan_dft_r2c_2d . . . . . . . . . . . . . . . . . . . fftw_mpi_plan_dft_r2c_3d . . . . . . . . . . . . . . . . . . . fftw_mpi_plan_many_dft . . . . . . . . . . . . . . . . . . . . . . 39 39 26 84 40 83 84 80 41 25 24 22 30 83 40 85 40 40 84 68 88 80 80 64 74 68 71 75 68 68 73 74 68 69 70 69 69 69 69 69 70 69 69 70 70 70 70 71 71 71 71 71 71 70

D
dfftw_destroy_plan . . . . . . . . . . . . . . . . . . . . . . . . . . dfftw_execute . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . dfftw_execute_dft . . . . . . . . . . . . . . . . . . . . . . . . 89, dfftw_execute_dft_r2c . . . . . . . . . . . . . . . . . . . . . . . dfftw_export_wisdom . . . . . . . . . . . . . . . . . . . . . . . . . dfftw_forget_wisdom . . . . . . . . . . . . . . . . . . . . . . . . . dfftw_import_system_wisdom . . . . . . . . . . . . . . . . . dfftw_import_wisdom . . . . . . . . . . . . . . . . . . . . . . . . . dfftw_init_threads . . . . . . . . . . . . . . . . . . . . . . . . . . dfftw_plan_dft_1d . . . . . . . . . . . . . . . . . . . . . . . . . . . . dfftw_plan_dft_3d . . . . . . . . . . . . . . . . . . . . . . . . . . . . dfftw_plan_dft_r2c_1d . . . . . . . . . . . . . . . . . . . . . . . dfftw_plan_dft_r2c_2d . . . . . . . . . . . . . . . . . . . . . . . dfftw_plan_with_nthreads . . . . . . . . . . . . . . . . . . . 90 88 90 90 91 91 91 91 90 90 90 90 91 90

F
fftw_alloc_complex . . . . . . . . . 3, 15, 22, 57, 79, 82 fftw_alloc_real . . . . . . . . . . . . 15, 22, 61, 62, 80, 82 FFTW_BACKWARD . . . . . . . . . . . . . . . . . . . . . . . . . . . 4, 7, 25 fftw_cleanup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23, 68 fftw_cleanup_threads . . . . . . . . . . . . . . . . . . . . . . . . 50 fftw_complex . . . . . . . . . . . . . . . . . . . . . . . . 4, 21, 77, 80 fftw_cost . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 FFTW_DESTROY_INPUT . . . . . . . . . . . . . . . . . . . 26, 66, 80 fftw_destroy_plan . . . . . . . . . . . . . . . . . . 4, 23, 66, 77 FFTW_DHT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13, 30 FFTW_ESTIMATE . . . . . . . . . . . . . . . . . . . . . 4, 18, 25, 100 fftw_execute . . . . . . . . . . . . . 4, 22, 38, 63, 66, 78, 81 fftw_execute_dft . . . . . . . . . . . . . . . . . . 39, 75, 77, 81 fftw_execute_dft_c2r . . . . . . . . . . . . . . . . . . . . 39, 81 fftw_execute_dft_r2c . . . . . . . . . . . . . . . . . 39, 79, 81 fftw_execute_r2r . . . . . . . . . . . . . . . . . . . . . . . . . 39, 81 fftw_execute_split_dft . . . . . . . . . . . . . . . . . . . . . . 39

110

FFTW 3.3.3

fftw_mpi_plan_many_dft_c2r . . . . . . . . . . . . . . . . . fftw_mpi_plan_many_dft_r2c . . . . . . . . . . . . . . . . . fftw_mpi_plan_many_transpose . . . . . . . . . . . 63, fftw_mpi_plan_transpose . . . . . . . . . . . . . . . . . 63, FFTW_MPI_SCRAMBLED_IN . . . . . . . . . . . . . . . . 60, 70, FFTW_MPI_SCRAMBLED_OUT . . . . . . . . . . . . . . . 60, 70, FFTW_MPI_TRANSPOSED_IN . . . . . . . . . . . . . . . 58, 63, FFTW_MPI_TRANSPOSED_OUT . . . . . . . . . . . . . 58, 63, FFTW_NO_TIMELIMIT . . . . . . . . . . . . . . . . . . . . . . . . . . . . FFTW_PATIENT . . . . . . . . . . . . . . . . . . . . 5, 18, 26, 51, fftw_plan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4, 22, fftw_plan_dft . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5, fftw_plan_dft_1d . . . . . . . . . . . . . . . . . . . . . . . . . . 4, fftw_plan_dft_2d . . . . . . . . . . . . . . . . . . . . . . . 5, 24, fftw_plan_dft_3d . . . . . . . . . . . . . . . . . . . . . . . 5, 24, fftw_plan_dft_c2r . . . . . . . . . . . . . . . . . . . . . . . . . . . . fftw_plan_dft_c2r_1d . . . . . . . . . . . . . . . . . . . . . . 7, fftw_plan_dft_c2r_2d . . . . . . . . . . . . . . . . . . . . . . . . fftw_plan_dft_c2r_3d . . . . . . . . . . . . . . . . . . . . . . . . fftw_plan_dft_r2c . . . . . . . . . . . . . . . . . . . . . . . . . 7, fftw_plan_dft_r2c_1d . . . . . . . . . . . . . . . . . . . . . . 7, fftw_plan_dft_r2c_2d . . . . . . . . . . . . . . . . . . . . . . 7, fftw_plan_dft_r2c_3d . . . . . . . . . . . . . . . . . . 7, 27, fftw_plan_guru_dft . . . . . . . . . . . . . . . . . . . . . . . . . . fftw_plan_guru_dft_c2r . . . . . . . . . . . . . . . . . . . . . . fftw_plan_guru_dft_r2c . . . . . . . . . . . . . . . . . . . . . . fftw_plan_guru_r2r . . . . . . . . . . . . . . . . . . . . . . . . . . fftw_plan_guru_split_dft . . . . . . . . . . . . . . . . . . . fftw_plan_guru_split_dft_c2r . . . . . . . . . . . . . . . fftw_plan_guru_split_dft_r2c . . . . . . . . . . . . . . . fftw_plan_guru64_dft . . . . . . . . . . . . . . . . . . . . . . . . fftw_plan_many_dft . . . . . . . . . . . . . . . . . . . . . . . . . . fftw_plan_many_dft_c2r . . . . . . . . . . . . . . . . . . . . . . fftw_plan_many_dft_r2c . . . . . . . . . . . . . . . . . . . . . . fftw_plan_many_r2r . . . . . . . . . . . . . . . . . . . . . . . . . . fftw_plan_r2r . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10, fftw_plan_r2r_1d . . . . . . . . . . . . . . . . . . . . . . . . . 10, fftw_plan_r2r_2d . . . . . . . . . . . . . . . . . . . . . . . . . 10, fftw_plan_r2r_3d . . . . . . . . . . . . . . . . . . . . . . . . . 10, fftw_plan_with_nthreads . . . . . . . . . . . . . . . . . 50,

72 72 73 73 71 71 71 71 26 64 80 24 24 77 79 28 28 28 28 27 27 27 79 35 36 36 37 35 36 36 38 31 33 33 33 29 29 29 29 67

FFTW_PRESERVE_INPUT . . . . . . . . . . . . . . . . . . . . . . . 7, fftw_print_plan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . FFTW_R2HC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11, fftw_r2r_kind . . . . . . . . . . . . . . . . . . . . . . . . . 10, 62, FFTW_REDFT00 . . . . . . . . . . . . . . . . . . . . . . . . . . 12, 29, FFTW_REDFT01 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12, FFTW_REDFT10 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12, FFTW_REDFT11 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12, FFTW_RODFT00 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12, FFTW_RODFT01 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12, FFTW_RODFT10 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12, FFTW_RODFT11 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12, fftw_set_timelimit . . . . . . . . . . . . . . . . . . . . . . . . . . FFTW_TRANSPOSED_IN . . . . . . . . . . . . . . . . . . . . . . . . . . FFTW_TRANSPOSED_OUT . . . . . . . . . . . . . . . . . . . . . . . . . FFTW_UNALIGNED . . . . . . . . . . . . . . . . . . . . 26, 39, 82, FFTW_WISDOM_ONLY . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

26 24 30 80 30 31 31 31 31 31 31 31 26 61 61 89 26

M
MPI_Alltoall . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . MPI_Barrier . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . MPI_COMM_WORLD . . . . . . . . . . . . . . . . . . . . . . . . . . . 53, MPI_Init . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64 65 55 55

P
ptrdiff_t . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38, 55, 80

R
R2HC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . REDFT00 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . REDFT01 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . REDFT10 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . REDFT11 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . RODFT00 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44, RODFT01 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . RODFT10 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . RODFT11 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42 43 44 43 44 45 45 45 45